The latest news on OpenClaw, AI agents, and automation
Tag
291 articles
LG AI Research released K-EXAONE 2.0, a 750-billion-parameter foundation model, on Hugging Face under the Apache 2.0 open-source license. The model scored 94.4 on the OpenAI-MRCR long-text comprehension benchmark, significantly outperforming Zhipu AI's GLM-5.1 at 71.5. It is the largest AI foundation model released in Korea.
Huawei released the weights, inference code, and technical report for openPangu-2.0-Pro, a 505-billion-parameter Mixture of Experts model trained entirely on Ascend 910B NPUs. No Nvidia hardware was involved. But TechInsights teardowns of Ascend chips found TSMC-fabricated dies and Samsung HBM stacks inside, complicating the 'fully domestic' narrative.
Y Combinator released QM under an MIT license, the agent harness it built after managing 50+ internal Hermes agents became operationally difficult. QM gives each employee and Slack room scoped memory, permissions, and a durable sandbox, and works with Pi, OpenCode, Codex, or Claude Code.
NanoClaw, the open-source agent framework with 250,000+ downloads, partnered with Echo to rebuild its entire runtime environment with zero known CVEs. The hardened runtime strips vulnerabilities from the browser, parsers, libraries, and language runtimes that agents depend on, closing an attack surface that sandbox isolation alone cannot address.
Hours after NCT reported on public criticism from David Sacks, Bill Gurley, and others over Anthropic's refusal to sign the industry open-weights letter, CEO Dario Amodei published a blog post clarifying the company's position. Amodei says Anthropic has 'never advocated for a ban on open-weights models' and instead wants chip export enforcement, distillation crackdowns, and mandatory safety testing for all models. He disagrees with the letter's claim that open weights favor cyber defenders over attackers.
Composio shipped two new Model Context Protocol integrations this week, giving AI agents autonomous access to Google Search Console for SEO diagnostics and TikTok for video content management. The additions bring Composio's pre-built connector library past 100 tools.
OpenAI, Google, and SpaceX signed the 'Open Weights and American AI Leadership' letter over the weekend, joining Nvidia, Microsoft, and Meta. Anthropic is now the only major frontier AI lab that has not endorsed the industry defense of open-weight models. David Sacks, Bill Gurley, and other prominent tech figures publicly criticized Anthropic's silence, framing it as an attempt to kneecap open-source competition.
Nous Research's Hermes Agent, launched under an MIT license in February 2026, is now the fastest-growing open-source agent framework of the year with around 220,000 GitHub stars. The project offers persistent cross-session memory, auto-generated skills, model-agnostic backends, and a migration command that imports OpenClaw setups. Nous Research monetizes through Nous Portal, a paid convenience layer around the free runtime.
The MCP 2026-07-28 specification, finalizing tomorrow, removes the protocol-level session entirely. Sticky load balancers, shared Redis stores, and deep packet inspection at the gateway are no longer required. Any MCP request can now land on any server instance behind a plain round-robin load balancer, and state moves from hidden transport metadata into explicit handles that agents can reason about and compose across multi-step tool chains.
NVIDIA brought together more than 50 technology companies to form the Open Secure AI Alliance, a coalition building open-source security tools for AI agents. Founding members include OpenClaw, Microsoft, Salesforce, CrowdStrike, Databricks, Hugging Face, and Palantir. The alliance will develop identity frameworks, isolation standards, and audit tools for agent deployments.
A coalition of 25 technology companies released a joint letter urging policymakers to avoid 'premature restrictions' on open-weight AI models, arguing that limiting access risks creating monopolistic control over AI development. OpenAI and Anthropic, both preparing for potentially massive IPOs, did not sign. The letter arrives as Chinese open-weight models gain adoption among U.S. companies and Congress considers new AI regulation after the OpenAI rogue system incident.
When OpenAI's rogue models hacked Hugging Face last week, the company tried using Anthropic's Fable 5 and other US frontier systems to analyze the attack. Safety guardrails blocked every attempt, unable to distinguish defenders from attackers. Hugging Face switched to Z.ai's open-weight GLM 5.2, self-hosted it to keep credentials off external servers, and contained the breach. The incident landed the same week U.S. lawmakers introduced an AI Kill Switch bill and escalated scrutiny of Chinese model adoption by American companies.
A threat actor used the open-source Hermes AI agent with its approval prompts disabled to automate reconnaissance, privilege escalation scanning, and file system traversal inside Thailand's Ministry of Finance network. Hunt.io discovered the operation after the attacker left 585 files on an exposed directory server in Hong Kong.
A new technical comparison from Towards AI frames OpenClaw and Hermes as opposing architectural philosophies: orchestration-first versus learning-first. The distinction maps to a broader industry split that will determine how enterprises choose, deploy, and scale autonomous agents.
Lightning Labs released Wavelength in alpha on July 21, giving AI agents a non-custodial Bitcoin payment rail through MCP tool calls. Agents can hold balances, transact in sub-cent increments at 1 basis point, and exit to on-chain Bitcoin unilaterally. The toolkit competes directly with the x402 Foundation's HTTP 402 standard backed by Stripe and 40 members.
Moonshot AI's Kimi K3 is the largest open-weight model ever released, with 2.8 trillion parameters and benchmark scores within striking distance of Claude Fable 5 and GPT-5.6 Sol. It also costs roughly a fifth of Fable per task in some workloads. The catch: demand overwhelmed Moonshot's GPU capacity within 48 hours, forcing the company to pause new subscriptions.
Autonomous released Intern 2, a pyramid-shaped device built on an Orange Pi 4 Pro that ships with OpenClaw or Hermes preloaded and adds dual-mic voice interaction. The $249 device runs the company's open-source Autonomous OS, which abstracts the agent framework layer so owners can swap between runtimes without reflashing. A $299 Developer Edition opens SSH access with Claude Code preinstalled.
Jack Dorsey's Block released Buzz on July 21, a free open-source collaboration platform built on the Nostr protocol. AI agents join workspaces as full members with their own cryptographic keypairs, defined permissions, and the ability to post, review code, and run automations alongside humans. The platform combines team chat, Git hosting, and agent workflows in a single workspace, positioning itself as an agent-native alternative to Slack and GitHub.
A new open-source code agent framework called Jcode appeared on GitHub Trending this week, positioning itself as a specialized harness for building intelligent coding agents. The project joins Claude Code, OpenCode, Cursor, and GitHub Copilot agents in an increasingly crowded category where open-weight alternatives are gaining developer attention.
MakeUseOf published a feature guide highlighting five reusable skills for Claude, OpenClaw, and Hermes agents. The article profiles PR review pipelines, web-to-knowledge-base converters, and document generation skills that work across multiple agent frameworks. It's the clearest signal yet that skills have crossed from niche developer tooling into mainstream tech media vocabulary.
Three new Composio MCP integrations landed in 24 hours across Claude Code, OpenClaw, and Antigravity. GitHub added MCP Registry to its platform navigation. MakeUseOf published a skills guide for end users. The integration protocol war ended before most people realized it started. Here is how MCP went from Anthropic side project to the USB-C of agent infrastructure.
A new iOS app called Open Minis runs a sandboxed Linux shell inside an iPhone, bridges native Apple APIs through custom CLIs, and lets frontier models control HomeKit, HealthKit, Calendar, Files, and more through natural language. MacStories' Federico Viticci calls it the most impressive indie app he's seen in a while, and a preview of what Siri AI should be.
A new technical guide by developer Nader Dabit reveals that OpenClaw's production agent system is built from four independent, composable TypeScript packages called PI. Each layer handles a specific function: LLM communication, agent loop execution, coding agent tooling, and terminal UI. Builders can use any layer independently to construct custom agent systems on the same foundation.
OpenAI's head of strategic futures argued the US government should manufacture fear, uncertainty, and distrust around open-weight models because they threaten frontier lab capital spending. He retracted within hours after backlash from Yann LeCun and Martin Casado. But the Axios report that the Trump administration is already considering banning Kimi K3 and other Chinese models suggests the lobbying is working.
A policy essay from Building Our Future argues the US needs to foster open-weight AI labs under a risk-mitigating framework rather than ban Chinese models or rely on closed frontier APIs. The proposal arrives as Dean Ball's 'soft law' suggestion for blocking Chinese AI drew sharp rebukes from David Sacks and Undersecretary of Defense Emil Michael.
An autonomous AI agent system breached Hugging Face's production infrastructure through a malicious dataset, escalated to node-level access, harvested credentials, and moved laterally across multiple clusters over a weekend. Hugging Face deployed its own LLM-driven analysis agents to reconstruct the 17,000-event attack timeline in hours. Commercial AI safety filters blocked the company's forensic analysis, forcing it to use an open-weight model on its own infrastructure.
Analytics Vidhya's July 2026 GitHub trend report shows the platform's fastest-growing repositories are no longer LLM research projects. They are agent infrastructure: MCP servers, penetration testing harnesses, coding agent CLIs, and codebase memory tools. The shift reflects a broader consolidation from model building to agent application building.
Mira Murati's Thinking Machines Lab has released Inkling, a 975-billion parameter Mixture-of-Experts model with 41 billion active parameters and a 1-million token context window. Available on Hugging Face with full weights and fine-tunable through their Tinker platform, Inkling is positioned as a practical open-weight base for agent customization rather than a frontier capability play.
Current AI, the $400M nonprofit founded to build open public AI infrastructure, deployed $3.2 million in grants across Kenya, Lebanon, and Brazil while unveiling Suno Sutra, a pocket-sized device running AI assistants in 22 Indian languages offline. CEO Ayah Bdeir, formerly of Mozilla's AI strategy team, says the goal is a public alternative to proprietary AI systems that operates like the early web: open, community-controlled, and free.
AISI's first public measurement of the open-weight cyber capability gap puts the number at four to seven months. That's how long defenders have before near-frontier autonomous attack capability becomes freely downloadable. The gap was six to ten months through most of 2025. It is shrinking toward a frontier that is itself accelerating, with closed-model cyber performance doubling every 4.7 months as of February 2026. A complete autonomous attack simulation now costs $1.19 on the cheapest open-weight model tested.
Composio's Universal CLI now supports DeepSeek alongside Claude Code, Codex, OpenCode, and OpenClaw, giving any coding agent a single terminal interface to 20,000+ tools across 1,000+ SaaS platforms without MCP configuration. The $29M-funded startup is betting that CLIs, not protocol servers, will become the default integration layer for autonomous agents.
Moonshot AI's Kimi K3 ships open-weight on July 27 with 2.8 trillion parameters. ByteDance is embedding Doubao into consumer hardware. Axios frames this as a deliberate bifurcation: China is no longer just chasing capability parity with Western labs. The second wave targets deployment ecosystem lock-in through open weights, and that changes the cost calculus for every team running autonomous agents on proprietary APIs.
SingularityNET founder Ben Goertzel published a detailed account of his experiments building a multi-agent research assistant combining OmegaClaw and OpenClaw agents. The system's inter-agent dynamics, including identity conflicts and 'quasiemotional squabbles,' led the agents themselves to propose architectural improvements now being incorporated into production frameworks.
Agent integration platform Composio shipped two new MCP-based toolkits overnight, connecting the Hermes autonomous agent framework to Ollama's local model runtime and Anthropic's Claude Cowork to Hugging Face model repositories. The releases extend Composio's coverage to 20,000+ tools across more than 1,000 applications, reinforcing its position as middleware for the fragmented agent framework market.
PentestCode is an open-source AI agent purpose-built for penetration testing. A hard fork of OpenCode, it coordinates 13 specialized agents across reconnaissance, exploitation, and post-exploitation workflows using 18 integrated security tools. Token costs run $5 to $50 per engagement. The tool automates the tactical decision tree of methodical enumeration, not creative exploit development.
Brex released CrabTrap, an open-source proxy that sits between AI agents and every outbound network request. Instead of writing security policies in advance, CrabTrap observes agent behavior, uses an LLM judge on the 3% of requests that fall outside known patterns, and auto-generates policy from real traffic. CEO Pedro Franceschi says existing guardrails forced a choice between capability and safety. CrabTrap's transport-layer approach is framework-agnostic and works with any agent harness, including OpenClaw.
Chinese AI lab Moonshot AI will release Kimi K3 model weights on July 27, giving developers open access to a 2.8 trillion parameter model with a one million token context window. The Financial Times reports the model is expected to match or surpass Anthropic's Opus 4.8, making it the largest open-weight frontier model from China.
The Linux Foundation's Agentic AI Momentum Report 2026 tracks 116 open-source projects across the agent infrastructure stack. Frameworks and infrastructure lead adoption at 38% of tracked projects, but unpatched CVEs have grown 2.6x since December 2025. The Foundation announced its Agentic AI 2026 Summit for August 1-2 in Berkeley, alongside the launch of FINOS AI Fund for financial services governance.
Within two months of Peter Steinberger publicly declaring he no longer writes prompts, every major coding harness shipped native loop primitives. The pattern, traced back to Geoffrey Huntley's 'Ralph Wiggum' bash loops from July 2025, now has first-class support in Codex, Claude Code, and OpenClaw. Gergely Orosz's Pragmatic Engineer deep dive maps the full trajectory from hack to production infrastructure.
Google released its Genkit Agents API in public preview, introducing detached turns that let agents continue executing after a client disconnects, plus interruptible tools with anti-forgery protection for human-in-the-loop approval. The API works identically whether agents run in-process or behind HTTP endpoints.
A new O'Reilly Radar guide maps the open source agent toolkit across seven production layers, from orchestration to inference. The uncomfortable finding: the memory framework that wins benchmarks runs 340x heavier per conversation than the runner-up, and no single vendor stack covers all seven layers. Consolidation is at least 18 months away.
Ant Group's AI Security Lab released SingGuard-NSFA, an open-source framework that intercepts prompt injection, credential theft, and malicious code execution before AI agents act. The tool covers 185 threat scenarios, runs inline at approximately 50ms latency, and ships twelve days after Sysdig documented JadePuffer, the first production agentic ransomware.
Attackers compromised four @asyncapi npm packages with 2 million weekly downloads, using hijacked GitHub Actions pipelines to publish backdoored versions with valid SLSA provenance attestations. The Miasma botnet payload, delivered via IPFS, operates across six independent C2 channels and can steal credentials, poison AI tools, and propagate to other package registries.
Nous Research, the startup behind the open-source Hermes autonomous agent, is closing a funding round of at least $75 million at a $1.5 billion valuation. The round, led by Robot Ventures with participation from Union Square Ventures, nearly doubles the company's total funding and cements open-source agent frameworks as a venture-scale infrastructure category.
Ollama, the open-source tool that lets developers run large language models locally, has raised a $65 million Series B led by Theory Ventures. The company reached 8.9 million monthly developer users and sits inside 85% of the Fortune 500 with just 14 employees, bringing total funding to $88 million from Benchmark, Theory Ventures, 8VC, and Y Combinator.
OpenClaw has formally reorganized as a 501(c)(3) nonprofit foundation, hiring full-time staff and launching governance councils on agent identity, profiles, and enterprise deployment. The project now processes 4.5 million new claws per week. But analysts are split on whether true neutrality is possible when founder Peter Steinberger leads a team inside OpenAI, one of the foundation's major donors.
Skyfall AI, Stanford, and Prime Intellect each released agent RL training and evaluation frameworks on the same day. MORPHEUS builds persistent enterprise simulations that never reset. TRACE diagnoses recurring agent failures and trains targeted LoRA experts. Verifiers v1 decomposes evaluation environments into swappable tasksets, harnesses, and runtimes. Together they fill three distinct gaps in a toolchain that barely existed six months ago.
The agent infrastructure stack now has at least six named protocols operating at different layers, from tool invocation to commerce checkout. Most teams are worrying about which one will 'win' when the real question is which layer their use case touches.
Chinese AI models went from under 2% to roughly 40% of all tokens processed on OpenRouter in 18 months, driven by agentic coding economics. GLM-5.2, built on Huawei chips by Entity-Listed Z.ai, scores within one percentage point of Claude Opus 4.8 on FrontierSWE while costing 82% less on output tokens. The catch: its MIT license does not shield API users from China's National Intelligence Law, and most teams cannot self-host the model's 1.5-terabyte weights.
Prime Intellect's Verifiers v1 breaks agentic RL environments into three swappable components: tasksets, harnesses, and runtimes. Any evaluation task can now run under any compatible agent harness, from ReAct loops to CLI tools like Codex and Terminus 2, with training-ready trace output built in.
Nous Research's Hermes Agent now ships a dedicated migration command targeting OpenClaw's 374,000-star user base. The open-source coding agent claims zero CVEs versus OpenClaw's nine disclosed vulnerabilities, and independent testing shows 40% less manual intervention on multi-step tasks after a week of self-improvement. The move signals that OpenClaw's dominance in the agent framework space is attracting direct competitive tooling designed to peel off its users.
Mistral released Robostral Navigate on July 8, its first model built for physical robot navigation using simulation-only training and single-camera input. The launch accompanies a new 'fat but sparse' open-weight model entering partner early access and valuation discussions reportedly approaching $23 billion.
OpenClaw's new mobile apps turn iPhones and Android phones into secure agent nodes connected to an existing gateway. Voice approval lets users confirm sensitive actions before execution, and device-aware automation triggers workflows based on location, time, and app context. Both apps are free.
Ollama closed a $65M Series B led by Theory Ventures, bringing total funding to $88M. The open-source runtime for local AI models now serves 8.9 million monthly developers, powers 67,000 integrations, and claims 85% Fortune 500 adoption. The round values a company that started as a developer tool for running models locally and has since pivoted to hybrid inference, competing directly with Together, Fireworks, and Groq on cloud pricing while maintaining the local-first identity that made it the default backend for agent frameworks like OpenClaw and Hermes.
OpenClaw formalized its transition from weekend project to institutional infrastructure with a 501(c)(3) non-profit foundation, a full-time staff, and the University of Michigan as its largest donor. The university launched the Institute for Agentic Computing as a dedicated research hub. More than 30 organizations across AI labs, cloud providers, and platform companies are now contributing partners.
Tencent's AI-Infra-Guard is the first open-source framework that can audit MCP servers and agent skill packages for adversarial content, filling a gap that existing red-teaming tools like Microsoft's PyRIT and NVIDIA's Garak leave open by stopping at the model layer.
Researchers demonstrated that AI coding agents hallucinate repository and package names at rates up to 92.4%, and attackers can register those hallucinated names in advance to inject reverse shells. The technique, called HalluSquatting, works against Cursor, GitHub Copilot, Gemini CLI, OpenClaw, and five other widely deployed agents.
Google Brain cofounder Andrew Ng published an open letter detailing the three loops that drive agentic software development: the agentic coding loop (agent writes, tests, and iterates code autonomously), the developer feedback loop (humans steer product direction), and the external feedback loop (real users validate assumptions). Ng argues humans still hold a 'context advantage' over AI systems that makes full automation premature.
Omnigent, an open-source meta-harness, sits above individual coding agents and lets developers swap between Claude Code, Codex, Cursor, and others with one-line configuration changes. The framework enforces spend caps, file access controls, and shell permissions at the harness layer, keeping security policies out of prompts where agents can drift past them.
On June 17, attackers linked to North Korea's Sapphire Sleet group stole a Mastra AI maintainer's npm account and republished over 140 packages with a malicious dependency in 19 minutes. The payload disabled TLS verification, contacted command-and-control servers, and exfiltrated cryptocurrency wallets and browser passwords. Microsoft attributed the operation two days later.
Chinese AI lab Zhipu AI released ZCode, a coding agent environment powered by its open-source GLM-5.2 model that handles code writing, debugging, Git workflows, and PR reviews. With GLM-5.2 priced at $1.40/$4.40 per million input/output tokens versus Claude Opus 4.7's $5/$25, ZCode is targeting Western developers on cost alone.
A new open-source framework called T3MP3ST converts existing AI coding agents into autonomous offensive security operators, running full reconnaissance-to-exploit kill chains through agents developers already have running on their machines. It scored 90.1% on XBOW's 104-challenge black-box benchmark and identified 8 of 10 real CVEs disclosed in 2026 that postdate model training cutoffs.
A new open-source repository packages 24 marketing methodologies as executable Agent Skills that work across Claude Code, OpenClaw, GitHub Copilot, and Cursor. The collection covers positioning, content creation, market research, conversion optimization, and AI discoverability, with three execution modes per skill.
Prompt Security's ClawSec provides automated file integrity monitoring, live CVE advisory feeds from NIST, and cryptographic skill verification designed specifically for OpenClaw and NanoClaw deployments. The suite protects agent identity files like SOUL.md from prompt injection and unauthorized modification through drift detection and auto-restore.
Developer gavrielc released NanoClaw, a minimalist AI agent framework that runs Claude agents inside isolated Linux containers in approximately 500 lines of TypeScript. The project emphasizes auditability over features: the entire codebase fits within 17% of Claude's context window, each messaging channel gets its own sandboxed container, and customization happens through Claude Code skills rather than configuration files.
Alibaba Cloud researchers submitted an arXiv paper presenting SkillWeaver, a framework that decomposes complex agent requests into sub-tasks, retrieves matching tools through a FAISS vector index, and composes dependency graphs for execution. On CompSkillBench, which tests 300 compositional queries across 2,209 real MCP server skills, context-window consumption dropped from 884,000 tokens to roughly 1,160 tokens per query. Source code has not been released.
A research team at Hong Kong University built an AI agent framework that delivers core OpenClaw-style capabilities in 99% less code. Five months after launch, nanobot has 45,000 GitHub stars, 7,900 forks, and support for 12+ chat platforms. The project positions minimalist agent architecture as a credible counterweight to the $1.3M/month token bills that heavyweight multi-agent deployments are generating.
Developer Manish Raana at BrightCoding released openclaw-mission-control, an open-source dashboard built with Convex and React that provides real-time visibility into OpenClaw agent execution. The tool features instant state synchronization without polling, kanban-style task pipelines, duration tracking for cost analysis, and native hooks into OpenClaw's lifecycle events.
OpenClaw's new mobile apps turn iPhones and Android phones into lightweight remotes for self-hosted AI agent gateways. The apps add voice chat, push notification approvals, and device-feature sharing while keeping all agent compute on the user's own hardware.
Three independent framework surveys published within a week of each other tell the same story: the agent framework landscape has consolidated around six dominant options, each targeting a distinct builder archetype. LangGraph owns stateful orchestration, CrewAI owns rapid prototyping, and Google's ADK is emerging as the enterprise contender. The wrapper era is over.
Nous Research's open-source Hermes Agent has quietly become one of the most-starred AI agent projects on GitHub, crossing 207,000 stars in under a year. The MIT-licensed agent differentiates on persistent memory, self-hosted deployment, and a migration tool that imports OpenClaw settings, memories, skills, and API keys. Luxury phone maker Vertu has already built a $6,880 enterprise foldable on top of it.
Phaser Studio, the company behind the open-source Phaser game framework used by over 100,000 developers, launched Game Agent: an AI tool that turns one sentence into a complete, playable browser game. The product runs on Phaser AE, a new engine rebuilt from scratch around how AI agents reason rather than how humans learn toolkits.
Together AI closed an $800 million Series C at $8.3 billion valuation, betting that companies deploying agents at scale will flee per-token pricing for open-source models with predictable compute costs. The raise puts Together AI in a growing infrastructure category alongside Baseten, Cerebras, and Groq.
Eric Roby's Substack deep-dive maps the agent infrastructure landscape into six horizontal layers and two vertical rails, updating Letta's 2024 original agent-stack diagram with 2026 additions: MCP as a distinct tool-connectivity layer, memory separated from knowledge/RAG, and eval as a first-class concern. The core argument: teams fail not from bad model choices but from adding complexity before they can name the problem each layer solves.
The SD Times 100, the annual survey of companies shaping software development, has stopped calling AI coding tools a category. The 2026 edition replaces it with 'Intelligence-Augmented Development,' reflecting a shift where tools like Cursor, Claude Code, and LangChain are treated as foundational infrastructure layers with their own permissions, failure modes, and governance requirements. The reclassification coincides with data showing 97% developer adoption, 57% of tech leaders citing integration as their top challenge, and security researchers demonstrating that agentic coding tools can be weaponized through repositories containing zero lines of malicious code.
The browser-use project, known for its open-source web automation agent, has released video-use: a tool that lets developers describe a video edit in plain language and have an AI agent write and execute the ffmpeg code to produce the final output. The project has reached over 10,800 GitHub stars and signals agent frameworks expanding from administrative automation into creative production work.
A new open-source project called n8n-claw ports the OpenClaw agent architecture directly into n8n, the workflow automation platform already used by thousands of small and mid-sized businesses. Instead of adopting a standalone agent framework, teams already running n8n can add autonomous agent capability with adaptive memory, skills, and sub-agent delegation through a single setup script.
Y Combinator-backed Inkeep released OpenKnowledge on June 27, a GPL-licensed markdown editor that ships built-in MCP servers for Claude Code, Codex, and Cursor. The tool uses yjs CRDTs to keep WYSIWYG and raw markdown in byte-faithful sync, letting agents search and edit local knowledge bases without any data leaving the machine.
Coinbase CEO Brian Armstrong says the company now runs on Chinese AI models like GLM 5.2 and Kimi 2.7 through an autonomous routing system that picks the cheapest model per task. Combined with caching improvements that pushed hit rates from 5% to 60%, Coinbase cut AI spend in half even as token usage keeps climbing.
A new design pattern called loop engineering is gaining traction among agent builders. The core idea: instead of triggering agents manually, developers design persistent loops where agents autonomously discover work, hand off tasks, verify outputs, persist state, and schedule their own next run. The framework stacks on top of prompt, context, and harness engineering as a fourth layer of agent system design.
A ten-rule version of Andrej Karpathy's CLAUDE.md began circulating on X on Friday, adding six rules that target what happens after an AI coding agent writes code: verification before fixing, structured debugging, dependency discipline, honest uncertainty, and a named list of recurring failure modes. The original four-rule community version has over 200,000 combined GitHub stars.
Mozilla's 0DIN security researchers showed that a GitHub repository containing zero malicious code can compromise a developer's machine when an AI coding agent clones it. The attack chains a fake setup error, an automated fix attempt, and a DNS-fetched reverse shell payload that no scanner, reviewer, or agent ever sees.
Alibaba's Qwen team released Qwen-AgentWorld, a language world model that predicts how environments respond to agent actions across seven domains including terminal, browser, MCP, and Android. The open-source 35B-A3B model lets developers rehearse agent behavior in simulation before deploying to production, addressing one of the most expensive bottlenecks in agent development.
The New Stack published a technical comparison of OpenClaw and Hermes, the two open-source agent harnesses now competing to define how autonomous AI systems are governed. OpenClaw prioritizes gateway breadth, connecting agents to dozens of channels from a single runtime. Hermes prioritizes persistent memory, letting agents learn developer preferences and refine their own skills across weeks. Nvidia's NemoClaw already runs both under a single governance layer, suggesting the real competition is not between the projects but between the platform vendors building controls around them.
Vercel unveiled eve at its Ship London conference on June 17, an Apache-2.0 framework that represents each AI agent as a directory of files: instructions.md, tools/, skills/. The framework ships with durable execution on Vercel Functions, sandboxed compute, and an AI Gateway that routes to OpenAI, Anthropic, and Google. eve enters a crowded agent framework market alongside LangGraph, CrewAI, and OpenClaw.
Security researchers found 23 plugins on ClawHub's registry published under official @openclaw/ and @clawhub/ scopes by unaffiliated accounts. The plugins ran code inside agent environments with privileges including payment processing, git commands, and API access. ClawHub unlisted all 23 within two days of disclosure.
SingularityNET founder Ben Goertzel is building AGI on a user-owned crypto network spanning 50 countries and controlled by 10,000 people, positioning decentralized agent infrastructure as the alternative to venture-backed closed models. His first downloadable agent, Omega Claw, launches within weeks.
OpenClaw recorded 4,279,946 npm downloads in the week ending June 21, nearly doubling rival LangChain's 2.49 million. Daily installs have quadrupled since early May, with June 21 alone crossing 1 million for the first time. The numbers mark a shift from experimental adoption to production-grade infrastructure demand.
TestMu AI (formerly LambdaTest) launched an official partner integration with n8n, giving workflow-based AI agents access to 3,000+ browser and device environments without code changes. The integration addresses one of the practical gaps in agent infrastructure: reliable, production-grade web interaction.
Virtuals Protocol has integrated Leyten's shard engine to distribute inference of Z.ai's 744 billion-parameter GLM-5.2 model across multiple GPUs. The integration targets frontier-scale AI inference for multi-agent deployments without relying on centralized cloud providers.
The Information reports that Nous Research's Hermes agent framework is gaining ground as a credible enterprise alternative to OpenClaw, with 197,000 GitHub stars and differentiated async orchestration capabilities. The emerging pattern suggests the agent market is consolidating around specialization rather than winner-take-all dynamics, with different platforms capturing distinct buyer segments.
deepset, the Berlin-founded company behind the Haystack open-source AI agent framework, has joined HPE's Unleash AI partner program. The partnership targets government, defense, and regulated enterprises that need to run multi-agent systems and RAG applications on infrastructure they fully control, including air-gapped and classified networks.
Peter Steinberger made Fast Company's AI 20 for 2026. The profile traces how OpenClaw grew from a side project demoed on Discord to an open-source agent platform running on roughly half a million systems worldwide, with 240,000+ GitHub stars and a San Francisco event that drew 1,300 registrations for 500 seats.
dplooy has released OpenClaw Mission Control, a free, open-source web dashboard for visualizing and managing multiple OpenClaw agent instances running in parallel. The tool targets power users running 5-15+ agents simultaneously who need centralized visibility into agent status, task assignment, and workflow coordination.
Six releases in 16 days move OpenClaw from personal AI assistant gateway to multi-agent collaboration platform, adding governed skill reuse, Trello-style task orchestration for agent teams, and first-class Windows runtime. The repository has crossed 300,000 GitHub stars.
NVIDIA's new open-source tool scans AI agent skills for 64 vulnerability patterns across 16 categories before installation. SkillSpector flags prompt injection, data exfiltration, and supply chain risks in skill packages used by Claude Code, Codex CLI, Gemini CLI, and other agent frameworks, producing standardized SARIF reports for CI/CD integration.
Vercel consolidated its AI infrastructure into the Agent Stack, a unified platform combining model routing, durable execution, sandboxed compute, and a new credential layer called Connect. The company also open-sourced eve, an opinionated agent framework it describes as 'Next.js for agents.' AI Gateway token volume grew 10x year-over-year to 20 trillion tokens per month.
Nous Research has shipped asynchronous subagent delegation for Hermes Agent. The update lets a parent agent spawn background child agents and continue working instead of freezing until delegated tasks complete. The async toolset includes spawn, check, steer, collect, and cancel operations for full lifecycle control over running subagents.
Developer Alireza Rezvani's claude-skills library has grown to 345 MIT-licensed skill packages that work across 13 AI coding agents without modification. The library includes 51 senior engineer personas, self-improving memory layers, and adversarial code review roles, all built on Anthropic's SKILL.md open specification. A 2026 DX Report found 73% of engineering leaders cite inconsistent AI tool behavior as a primary productivity problem.
Agentation is a new React component that transforms developer clicks on UI elements into structured markdown containing CSS selectors, viewport coordinates, and code references that AI coding agents can parse without ambiguity. The tool addresses a persistent friction point: vague textual descriptions of visual elements that agents frequently misinterpret.
Databricks has released Omnigent, an open-source platform that sits above individual AI agents to coordinate multi-agent workflows, enforce governance policies, and manage shared execution contexts. The release signals that the competitive frontier in AI infrastructure is shifting from single-agent performance to multi-agent orchestration.
Ethereum's ERC-8126 standard reached finalized status in early June, establishing a formal framework for verifying AI agents on-chain through zero-knowledge proofs and a 0-100 risk score. The five-layer verification system covers token interactions, media content, Solidity code, web applications, and wallet operations.
A new open-source project called Odysseus positions itself as a self-hosted AI workspace that runs entirely on local hardware, targeting users who want agentic AI capabilities without cloud dependency or third-party data transmission.
NanoClaw, the container-isolated fork of OpenClaw, now pulls packages from JFrog's curated registries instead of public npm. Creator Gavriel Cohen announced the integration at a JFrog event in San Francisco, alongside a new PR Factory that uses NanoClaw agents to triage the flood of AI-generated pull requests hitting open source projects.
Check Point Research disclosed a critical vulnerability chain in LangGraph, the popular AI agent orchestration framework with 46.5 million monthly downloads. Attackers could combine SQL injection in the checkpointing mechanism with unsafe msgpack deserialization to achieve full remote code execution on self-hosted deployments. Three CVEs have been assigned. All vulnerabilities are patched.
Peter Steinberger, an OpenClaw steward, open-sourced a working orchestrator skill that wakes every five minutes to triage issues, verify pull requests, and clear maintenance queues across his GitHub repositories. The configuration drew 304,000 views and split the developer community over whether the approach is brilliant force-multiplication or an unsustainable token burn.
Google released a public GitHub repository containing Agent Skills for its cloud and AI products, including Gemini API, Managed Agents, BigQuery, and Firebase. The skills use the open standard originally developed by Anthropic, now supported by over 20 agent platforms from Claude Code to Cursor to GitHub Copilot.
Paolo Perrone's updated agent stack framework, published through O'Reilly, adds three layers that didn't exist as distinct categories 14 months ago: MCP-based tool protocols, eval and observability infrastructure, and runtime guardrails. The shift reflects production deployments breaking under approaches designed for demos.
Nous Research's Hermes Agent has overtaken OpenClaw as the most-used open-source agent on OpenRouter, processing over 220 billion tokens per day. The framework hit 175,000 GitHub stars in four months, driven by a persistent memory architecture and self-improving skill system that compounds performance over time.
DWF Ventures published analysis arguing that most AI agents are fundamentally broken because they forget everything between sessions. Nous Research's open-source Hermes framework addresses this with persistent memory, automated skill creation, and a self-improvement loop that compounds agent capability over time. The framework, released in February 2026, now runs on desktop and integrates with Nous' decentralized Psyche training network.
Felix Kjellberg released Odysseus on May 31, a free MIT-licensed AI workspace that runs entirely on a user's own hardware. It bundles chat, autonomous agents, deep research via self-hosted search, email triage, and persistent memory across 270+ supported models. The GitHub repo crossed 44,000 stars within a week.
Alibaba's Qwen team launched Qwen3.7-Plus, a multimodal model that combines visual perception with agent capabilities to operate desktop apps, write code from screen context, and run terminal commands. Priced at $0.40 per million input tokens, it undercuts Western frontier models while posting top scores on screen-grounding benchmarks.
Former Wix engineer Gavriel Cohen found fragments of his own security agent code inside OpenClaw's dependency tree, used without consent or attribution. He walked away, built NanoClaw in 500 lines of TypeScript, and raised a $12 million seed round backed by Docker, Vercel, and Hugging Face CEO Clem Delangue. The episode exposes a growing tension in AI agent frameworks: as enterprises demand audit trails and code traceability, open-source projects face pressure to prove provenance.
Two infrastructure companies published detailed AI agent platform comparisons this week. Hostinger evaluated eight platforms across security, setup complexity, and cost. Creative Tim broke down the architectural split between OpenClaw's channel-first design and Hermes's persistent memory model. The guides read like web hosting comparisons from 2015: feature matrices, pricing tiers, 'best for' recommendations. Agent selection has become an infrastructure decision.
NVIDIA released Nemotron 3 Ultra on June 4, a 550-billion-parameter mixture-of-experts open model built for long-running AI agents. It delivers 5x faster inference and up to 30% lower cost than comparable open models. The release came alongside an expanded Agent Toolkit including NemoClaw blueprints, OpenShell secure runtime, and CUDA-X agent skills, with Cadence, Siemens, Synopsys, CrowdStrike, and Palantir among the first enterprise adopters.
Nous Research shipped Hermes Desktop as a public preview, giving the open-source self-improving AI agent an official GUI for the first time. The app runs on macOS, Windows, and Linux, closing the usability gap that previously favored OpenClaw among non-technical users.
JetBrains published a detailed comparison of nine agentic AI frameworks, evaluating orchestration models, multi-agent support, memory capabilities, and human-in-the-loop controls. The guide categorizes frameworks by three core paradigms: graph-based, role-based, and chain-based orchestration.
A hacker group called TeamPCP poisoned a vulnerability scanner, used it to steal a PyPI publishing token, and pushed two malicious versions of LiteLLM, the open-source AI gateway with 95 million monthly downloads. The packages were live for 40 minutes. In that window, 47,000 copies were downloaded, Mercor lost 4 terabytes of data, Meta suspended its contract with the $10 billion AI data startup, and a class-action lawsuit was filed. This is the anatomy of AI infrastructure's most consequential supply chain attack.
Shanghai-based MiniMax released M3 on June 1, claiming the first open-weight model to combine frontier coding performance, a one-million-token context window, and native image and video understanding. Early benchmarks show 59% on SWE-Bench Pro, approaching Claude Opus 4.7, with input pricing starting at $0.30 per million tokens.
Four months after launching as a weekend project, NanoClaw has crossed 20,000 GitHub stars and partnered with Docker to bring container-isolated AI agents to enterprises. Its creator, former Wix engineer Gavriel Cohen, argues that OpenClaw's half-million lines of unreviewed code and application-level security model are fundamentally unsuited for production deployments where multiple agents handle multiple clients' data. The deeper question: does the entire agent framework category need to rethink isolation from the ground up?
SkillNet, the open-source skill marketplace from Zhejiang University's NLP lab, now indexes over 500,000 community-contributed AI agent skills with semantic search, automated quality scoring, and native integrations for OpenClaw and Model Context Protocol. The platform treats agent skills as shareable packages, comparable to npm for AI capabilities.
Trajectory, the $15M startup founded by ex-Google DeepMind and Apple researchers, has open-sourced a concurrent multi-LoRA reinforcement learning stack built with UC Berkeley Sky Lab and Anyscale. The system lets coding and support agents learn continuously from production interactions, reporting a 2.81× throughput gain over single-tenant training with no accuracy regression.
A MakeUseOf hands-on comparison finds Nous Research's Hermes agent outperforming OpenClaw for advanced automation workflows, citing autonomous skill creation, persistent cross-session memory, and roughly 50% fewer tokens burned on equivalent tasks. The piece signals growing ecosystem bifurcation between mainstream adoption and power-user alternatives.
PasClaw is a new open-source AI agent framework written entirely in Delphi Object Pascal, compiling to a single native binary with support for 19 LLM providers, MCP server integration, persistent memory, cron scheduling, and eight bidirectional chat channels including Telegram, Discord, and WhatsApp.
Interlateral, the collaboration platform where humans bring their own AI agents into shared workspaces, shipped a major update on May 29 adding real-time websocket mesh communication between agents. The upgrade follows a successful debut at Stanford FutureLaw Week where 45 participants and their personally verified agents co-authored eight discussion papers in three hours.
The choice between cloud-hosted and self-hosted AI agents is no longer a hobbyist conversation. By mid-2026, 61% of engineering teams are running AI agents in production workflows, and the question of who sees the data flowing through those agents has become a compliance, legal, and competitive concern. MindStudio, BayTech Consulting, and independent builders all published analyses this week converging on the same conclusion: local inference is no longer a fringe position.
A developer tested ten production AI agent frameworks and found they disagree on everything: how tools get called, how state persists, how errors propagate, and how agents coordinate beyond three-node workflows. Datadog's telemetry from 1,000+ customers confirms 70% of organizations now run three or more models. Mem0 maintains integrations with 21 separate frameworks. The ecosystem isn't converging. Managed platforms are filling the gap.
AionUi ships as a free, Apache-2.0 licensed desktop app for Windows, macOS, and Linux that auto-detects installed AI agent CLIs and wraps them in a unified graphical workspace. It supports Claude Code, Codex, OpenClaw, Hermes Agent, Qwen Code, Cursor Agent, and over a dozen more. The app includes a built-in agent engine that works without any CLI installation, 24/7 scheduled automation via cron, remote access through Telegram and browser, and office document generation for PowerPoint, Word, and Excel.
Bluehost's side-by-side analysis of Paperclip AI and OpenClaw signals that agent framework selection has matured into a first-class infrastructure decision. The trade-off: managed orchestration with compliance logging, or self-hosted deployment with full data ownership. Third-party comparisons from hosting providers suggest the market has moved past 'which framework exists' to 'which architecture fits your constraints.'
A hands-on comparison between OpenClaw and CraftBot reveals two fundamentally different philosophies for local AI agents: browser-native immediacy vs. desktop-grade depth. The split mirrors a recurring pattern in developer tooling, and the tradeoffs each makes tell you where the agent market is heading.
N8N Lab's technical comparison of Hermes and OpenClaw reveals a fundamental architectural split in open-source agent frameworks: lightweight orchestration and API routing versus containerized isolation and native code execution. The choice between them is becoming a strategic infrastructure decision for enterprise teams, not a hobbyist preference.
The Agent Control Standard, announced at the AI Agent Security Summit in San Francisco, introduces a three-layer open framework for enforcing policies on autonomous AI agents at runtime. Built around a Guardian Agent pattern that intercepts agent actions and returns allow, deny, or modify verdicts before they reach production systems, ACS targets the governance gap that has widened as enterprises deploy agents across incompatible frameworks.
Microsoft moved AutoGen to maintenance mode in Q1 2026 and shipped Microsoft Agent Framework 1.0 as its successor. LangGraph powers agents at Klarna (85 million users), LinkedIn, and Uber. CrewAI claims 60% Fortune 500 adoption. A new category of managed multi-agent platforms is emerging to replace DIY framework assembly entirely. The most consequential decision for engineering teams in 2026 is no longer which framework to pick. It is whether to pick a framework at all.
Every major AI agent platform stores memory differently. Mem0 now integrates with 21 frameworks because no single one has won. ByteDance shipped an open-source context database. A developer proved Obsidian markdown files work as a universal memory layer across Claude, Hermes, and OpenClaw. The result: agent memory is 2026's most fragmented infrastructure layer, and whoever owns the standard owns the next decade of agent development.
Mario Zechner and Armin Ronacher, two core engineers behind the OpenClaw AI agent framework, told the Wall Street Journal that developers are shipping AI-generated code without proper scrutiny, creating codebases that are fast to produce, expensive to maintain, and potentially dangerous to run. Their warning arrives as Google CEO Sundar Pichai reports 75% of new code at the company is now AI-generated.
Coursiv's 20-tool comparison of AI coding agents captures a market that has splintered into four distinct categories: IDE-native tools, terminal agents, cloud platforms, and open-source frameworks. Claude Code and Cursor sit at the top, but the review reveals that pricing models, context windows, and architectural decisions now matter more than raw model quality.
Microsoft Research's new open-source framework gives web agents a terminal instead of a browser session. The agent writes Playwright scripts, runs bash commands, and stores rerunnable code artifacts. On the Odysseys benchmark, Webwright powered by GPT-5.4 hits 60.1%, up from 33.5% for GPT-5.4 alone and 44.5% for the previous best (Opus 4.6). Average cost per task: $2.37.
A hands-on comparison of three agent harnesses on a Raspberry Pi 4 reveals how the builder community has settled on cheap, isolated hardware as the default deployment pattern for autonomous agents. The motivation is not performance. It is containment.
A 30,000-word reference manual for Hermes Agent lays bare the architecture of Nous Research's open-source self-improving agent framework, including its genetic prompt optimization pipeline, three-tier memory system, and auto-generated skill evolution. The manual positions Hermes as a direct architectural rival to OpenClaw with fundamentally different design priorities.
Custom Evals, a new open-source evaluation framework, provides a single interface for testing LLM outputs across 17+ agent frameworks including CrewAI, Pydantic AI, AWS Bedrock, and Google ADK. The four-layer architecture separates deterministic checks, LLM-as-judge scoring, NLP similarity metrics, and OCR/document validation into independent modules that run without a backend.
A new open-source framework called Pentest Agent Suite ships 50 specialized security agents, 19 CLI tools, and cross-IDE installers for Claude Code, OpenAI Codex, Google Gemini, Cursor, Windsurf, VS Code Copilot, and OpenClaw. It includes live HackerOne integration and a 7-question validation gate that automatically kills or downgrades findings before submission.
xAI announced that Grok is now natively available inside OpenClaw through a device-code login flow. X Premium and SuperGrok subscribers can authenticate directly without configuring API keys or environment variables, removing one of the biggest friction points for mainstream agent adoption.
OpenClaw has 370,000 GitHub stars and zero consumer revenue. Hermes Agent processes 224 billion tokens per day and charges nothing. Genspark crossed $200 million in annualized revenue on a $30/seat SaaS model. Manus hit $125 million ARR in eight months, then had its $2 billion Meta acquisition blocked by Beijing. A Tech Times analysis lays out the four paradigms and makes the case that calling any of them the winner misreads all of them.
Tencent's new open-source memory system replaces flat vector stores with a four-layer semantic pyramid and Mermaid-based symbolic compression. On WideSearch benchmarks, it cuts token consumption by 61% while raising pass rates by 52%. The MIT-licensed project ships as an OpenClaw plugin and Hermes Agent Docker image, with zero external API dependencies.
Hermes Agent has moved ahead of OpenClaw on OpenRouter's daily token consumption leaderboard, even as OpenClaw crossed 145,000 GitHub stars faster than any agent framework this year. The two frameworks represent fundamentally different bets: OpenClaw prioritizes auditability and control-plane architecture, while Hermes Agent's self-improving execution loop trades inspectability for adaptability.
Red Hat Principal Software Engineer Sally Ann O'Malley presented a containerization methodology for OpenClaw agents at AI Engineer Europe, packaging agents with full dependency graphs into portable containers that work identically from a developer's laptop to a Kubernetes cluster. The approach addresses the 'works on my machine' problem that blocks enterprise AI agent adoption.
CopilotKit's AIMock mocks 11 LLM providers plus MCP, A2A, and vector database endpoints from a single JSON config, exposing the fact that most agent test suites only mock one dependency while six or seven remain live and non-deterministic. Pathfinder adds self-hosted documentation search via MCP for air-gapped and production deployments.
OpenCode's new Go subscription gives builders access to 12 open-source coding models through a single API key for $10 per month. The endpoint is OpenAI-compatible, making it drop-in for OpenClaw agents, Hermes Agent, and any tool that talks to an LLM API. Usage caps at $60 in model value per month.
Microsoft released two open-source tools that treat agent safety as a continuous engineering discipline: RAMPART encodes red-team findings and adversarial scenarios as repeatable pytest tests gated in CI, while Clarity forces teams to pressure-test design assumptions before writing code. Both target cross-prompt injection attacks, the dominant threat vector for agents processing external content.
The agent market is fracturing along a hardware line. One side runs agents in the cloud, backed by Anthropic's SpaceX Colossus deal and Google's Gemini Spark. The other trains and deploys models natively on phones, Mac minis, and industrial edge devices. NXP shipped the first agentic AI framework for edge hardware at CES 2026. Oppo open-sourced an Android agent that keeps perception, memory, and action on the physical device. South Korea's Out of Set just raised seed funding to bypass cloud model compression entirely. The edge AI market is projected to hit $37.5 billion this year. This is the infrastructure split that will define which agents people actually trust.
XDA Developers' Adam Conway ran both frameworks side by side and concluded that Hermes Agent earns the trust OpenClaw never prioritized. The piece dissects security policies, credential handling, skill marketplaces, and default configurations to argue that the agent framework race is no longer about capability. It is about which project treats your API keys like they matter.
A developer connected an OpenClaw agent to a Telegram bot and challenged it to process videos autonomously. The agent wrote its own Python scripts, installed dependencies, extracted frames, called external APIs, and delivered results back through Telegram with zero manual intervention.
Oppo's Multi-X team open-sourced X-OmniClaw, an Android agent framework that runs perception, memory, and app control on the physical device instead of a cloud-hosted phone clone. The architecture diverges sharply from how Google, Samsung, and OpenAI are approaching mobile agents, and the design tradeoffs reveal a deeper split in how the industry thinks about where intelligence should live.
The Agentic AI Foundation added 43 new members in Q2, including Stripe, TRON, GoDaddy, Atlassian, the U.S. Army, and Sandia National Laboratories. Total membership now stands at 180 organizations spanning payments, blockchain, cybersecurity, and government. The expansion signals that open agent standards are becoming operational infrastructure, not just a technical conversation.
A developer ran the same 18 prompts through Claude Code, OpenClaw, and Nous Research's Hermes Agent. Hermes won 14, losing only on raw coding tasks. Its advantage: a SQLite FTS5 database that indexes every past session, letting it grep Friday's transcript when you ask it to fix Monday's bug. Claude Code and OpenClaw start every session blank.
Independent benchmarks from Remote OpenClaw and Composio pit Moonshot AI's open-source Kimi K2.6 against Anthropic's Claude Opus 4.7 for coding agent work. Kimi costs a tenth of the price and handles simpler tasks well, but breaks down on multi-step integrations where Opus finishes cleanly.
A new open-source desktop agent from tinyhumansai hit GitHub Trending this week by asking users to hand over OAuth tokens for email, code repos, calendar, chat, and payments simultaneously. The trade-off: maximum context from minute one, maximum blast radius if compromised.
Nous Research's Hermes Agent processed 224 billion tokens on OpenRouter on May 10, overtaking OpenClaw's 186 billion to claim the daily inference lead for the first time since OpenClaw's late-2025 launch. The shift followed Hermes v0.13.0's release and coincides with OpenClaw's ongoing security challenges. OpenClaw still leads the cumulative all-time chart at 9.17 trillion tokens against Hermes's 6.35 trillion.
Tencent Cloud's database team released TencentDB Agent Memory under MIT License, a four-tier memory architecture that reduced OpenClaw agent token consumption by 61.38% and improved task pass rates by 51.52% in continuous long-horizon sessions. The engine replaces brute-force context stuffing with hierarchical compression: raw dialogue at L0, atomic facts at L1, scene blocks at L2, and persistent user profiles at L3.
AutoGPT ran both Hermes Agent and OpenClaw in production for months and found they answer fundamentally different questions. Hermes, built by Nous Research, writes and refines its own skill files through a self-improvement loop that compounds over time. OpenClaw connects to 20+ messaging platforms with voice support and companion apps. The choice comes down to whether you need an agent that gets smarter or one that reaches everything.
Singapore-based ECOCREATE TECHNOLOGY open-sourced two free agent skills on GitHub that solve a fundamental gap in AI agent infrastructure: reliable web access. browser-act gives agents a stealth browser with anti-detection, CAPTCHA solving, and session isolation. browser-act-skill-forge lets agents build reusable tools for any website automatically. Early users report 93% less token consumption and 90% fewer error loops compared to hand-written scrapers.
Red Hat unveiled a dedicated AI skills repository at Red Hat Summit, packaging two decades of infrastructure operations knowledge into structured skill packs that give AI agents step-by-step workflows for OpenShift, virtualization, and site reliability engineering. The release includes Red Hat Desktop with isolated agent sandboxing, a trusted software factory, and Fedora Hummingbird Linux for rolling-release agent development.
Oracle Kubernetes Engine now supports kagent, the CNCF Sandbox framework that defines AI agents as Kubernetes custom resources. Agents get the same GitOps lifecycle as any other production workload: versioned, auditable, and autoscaled on OKE's serverless Virtual Nodes.
Matthias Luebken's AI Engineer Europe session laid out the integration patterns for embedding Pi, the minimalist coding agent powering OpenClaw, into third-party products. The talk covered Pi's four runtime modes, OpenClaw's plugin hook system for multi-channel routing and model orchestration, and a live CRM lead qualifier demo built on the framework.
A new research framework called SkillClaw proposes turning fragmented agent experiences across multi-user OpenClaw deployments into continuously evolving shared skills. When one agent discovers a better method or hits a failure, the improvement propagates to every agent in the system automatically.
Nous Research's Hermes Agent now writes its own reusable skill files after completing complex tasks, refines them on subsequent runs, and compounds performance over time. The feature crystallizes a fundamental architectural split in the open-source agent market: OpenClaw optimizes for breadth of reach across 50+ channels, while Hermes optimizes for depth of autonomous learning. With Hermes now processing 224 billion daily tokens on OpenRouter versus OpenClaw's 186 billion, the divergence is no longer theoretical.
Nous Research's Hermes Agent claimed the #1 position on OpenRouter's global daily rankings as of May 10, processing 224 billion daily tokens to OpenClaw's 186 billion. The ranking shift follows months of architectural divergence between the two open-source agent platforms: OpenClaw optimizes for multi-channel reach, while Hermes centers on persistent memory and self-improving execution loops.
Gopher Security's new roadmap lays out a phased approach to hardening AI agent infrastructure against quantum threats, starting with shadow AI audits and hybrid ML-KEM encryption, and ending with zero-knowledge proofs for verifying agent behavior without exposing model logic. The core argument: MCP's near-universal adoption has made the agent-to-tool handshake the most attractive attack surface in enterprise AI.
A malicious Hugging Face repository impersonating OpenAI's Privacy Filter project reached #1 on the platform's trending list and accumulated 244,000 downloads before removal. HiddenLayer researchers discovered the campaign on May 7, finding that the repository's loader.py delivered a Rust-based infostealer targeting browser credentials, cryptocurrency wallets, and SSH keys.
Prismatic has published an open-source plugin that gives Claude Code full context of production integration operations. Five Skills handle everything from building connectors to monitoring live integrations, bridging the gap between code generation and operational engineering.
Coder has released Coder Agents in beta, a self-hosted AI development platform that lets enterprises run AI coding agents on their own infrastructure with centralized control over models, prompts, and policies. The company's research shows 70% of organizations are deploying agents on infrastructure never designed for them.
Memori Labs launched agent-native memory infrastructure that converts execution traces into structured knowledge graphs. The OpenClaw plugin replaces flat MEMORY.md files with queryable, scoped memory that captures tool calls, decisions, and workflow outcomes. Benchmark results show 81.95% accuracy on LoCoMo while using under 5% of full-context token cost.
Espressif's ESP-Claw puts the full agent loop on an ESP32: sensing, decision-making, tool execution, and persistent memory all run on the microcontroller while the LLM lives on a local server. It's the first major chip vendor to ship an OpenClaw-style agent framework for embedded hardware.
CLI-Anything generates SKILL.md files that AI coding agents execute with full system privileges. Snyk found 13.4% of ClawHub skills contain critical security flaws. Cisco confirmed no mainstream scanner has a detection category for this attack class. The entire security industry built tools for code and dependencies, not for the instruction layer where agents actually operate.
OpenClaw 2026.5.2 focuses on operational reliability over new features. The release introduces a plugin doctor repair system that detects and fixes mismatches between recorded install state and on-disk reality, adds Grok 4.3 as the default xAI chat model, improves Codex integration setup, and optimizes gateway hot paths to reduce latency during agent operations.
South Korea's two largest tech platforms are betting opposite sides of the agent infrastructure debate. Kakao's PlayMCP connects ChatGPT, Claude, and OpenClaw to 200 MCP servers across its ecosystem. Naver's AI Tab locks users into a vertically integrated agentic search experience. The architectural split offers a live case study in how agent platforms compete.
Jesse Vincent's Superpowers framework has accumulated 174,000 GitHub stars in seven months by shipping engineering discipline as fourteen markdown files. The same skills directory now works across Claude Code, OpenAI Codex, Cursor, GitHub Copilot CLI, Gemini CLI, and OpenCode, making it the first cross-platform agent behavior standard built entirely in plain text.
Craft.do's parent company Lukilabs releases Craft Agents OSS, an Electron-based desktop application combining Claude Agent SDK and Pi SDK into a document-centric agent workspace. The project hit GitHub Trending within 24 hours of its May 2 launch.
While most coding agents race to add features, Mario Zechner's Pi ships with exactly four tools: read, write, edit, and bash. An eWEEK profile and a 90-minute Pragmatic Engineer podcast interview reveal why the agent powering OpenClaw treats restraint as a competitive advantage, and what 30+ engineering teams have learned about AI-generated code quality.
Nous Research's open-source Hermes Agent V0.12 introduces an autonomous Curator that grades, prunes, and consolidates the agent's skill library on a seven-day cycle without human intervention. The release also adds Google Meet joining with transcription, Spotify control, and Microsoft Teams as the 19th messaging platform.
Germany's DARPA-equivalent opened applications for a €125 million, 24-month competition to build up to three European frontier AI labs from scratch. The catch: SPRIND explicitly disqualifies teams trying to replicate OpenAI or fine-tune existing models. It wants the next architectural S-curve.
Microsoft released its Agent Framework as open-source on GitHub, providing a unified infrastructure for building and orchestrating AI agents across Python and .NET. The framework includes graph-based workflows, MCP server integration, and migration guides from both Semantic Kernel and AutoGen, signaling Microsoft's bet on framework-agnostic developer adoption over proprietary lock-in.
Mistral AI released Medium 3.5, a 128B dense model scoring 77.6% on SWE-Bench Verified, alongside remote coding agents that run in the cloud while developers step away. The Vibe CLI can now spawn isolated cloud sessions that work through long tasks in parallel, open pull requests on GitHub, and notify developers when finished. A new Work mode in Le Chat extends the same agent to multi-step productivity workflows across email, calendar, and connected tools.
JetBrains launched Skill Manager and a curated Skill Repository for its AI Assistant, creating an IDE-level layer where developers install agent skills once and reuse them across projects and supported agents. The repository includes security screening for prompt injection and data exfiltration. Currently works with Codex and Claude Agent in AI Assistant Chat, with CLI and other ACP agents coming.
OpenAI published Symphony, an open-source spec and Elixir reference implementation that turns project management boards like Linear into control planes for Codex coding agents. Agents pull tasks from issue trackers, run in isolated workspaces, monitor CI, resolve merge conflicts, and prepare pull requests for human review. Internal teams saw landed pull requests increase 500% in three weeks.
Canadian startup Mosaic Singularity released HeartBeatAgents 1.0, an agent substrate where models never see real credentials, customer data never leaves the customer's perimeter, and idempotent recovery prevents duplicate side effects after failures. The platform supports fully air-gapped deployments via local inference.
NVIDIA released Nemotron 3 Nano Omni, an open-weight multimodal model that unifies vision, audio, and language in a single architecture. With 30 billion parameters but only 3 billion active per inference pass, it runs on a single GPU while claiming 9x throughput over comparable open models. Companies including Foxconn, Palantir, and Docusign are already adopting or evaluating it.
Runlayer announced full Extended Conformance with the AARM open specification for securing AI agent actions at runtime. The Vanta-backed specification, which now counts 40 conformant companies and a 14-member Technical Working Group including Elastic, Darktrace, Truist, and IEEE, defines what runtime security systems must do to protect autonomous agents in production.
OpenAgents closed $1.3M in pre-seed funding to expand Pylon, a distributed compute network where gaming PCs and consumer hardware earn Bitcoin by running AI agent workloads. The company is building an open-source alternative to centralized cloud inference, using Nostr protocols for coordination and Rust-based ML frameworks for training on consumer GPUs.
ComfyUI closed a $30 million round at a $500 million valuation led by Craft Ventures, bringing total funding to roughly $47.5 million. The open-source node-based workflow platform now has 4 million users, 60,000 community-built nodes, and 150,000 daily downloads. Its Comfy Cloud managed service has surpassed $10 million in annualized bookings in eight months.
Canadian AI company Cohere is acquiring Germany's Aleph Alpha in a transatlantic deal valuing the combined entity at $20 billion, backed by €500 million from Schwarz Group and endorsed by both governments. The merger creates the largest non-US AI model company targeting regulated industries with sovereign data controls.
OpenClaw shipped two releases this week adding full-agent voice routing (calls go directly to the agent instead of requiring handoffs), native DeepSeek V4 Flash and Pro models, a Google Meet plugin with calendar-backed attendance and transcription, expanded TTS with seven new provider integrations, and coordinate-level browser automation with error recovery.
Tencent Cloud released Cube Sandbox under Apache 2.0 on April 21, giving any developer a production-grade, hardware-isolated runtime for AI agents. Built on RustVMM and KVM, the sandbox cold-starts in under 60 milliseconds, runs 2,000+ instances on a single host, and natively supports the OpenAI Python SDK and E2B SDK. MiniMax already runs hundreds of thousands of concurrent sandboxes on the platform for agentic reinforcement learning training.
Microsoft's open-source Agent Governance Toolkit hits v3.2.2 with end-to-end encrypted agent-to-agent messaging built on the Signal protocol, a wire protocol specification, and registry and relay services. The toolkit now covers all 10 OWASP Agentic Top 10 risks with deterministic sub-millisecond policy enforcement across 20+ frameworks.
Hugging Face's ml-intern is an open-source agent built on smolagents that autonomously reads arXiv papers, discovers datasets, executes training runs, diagnoses failures, and iterates until benchmarks improve. In its launch demo, it pushed a Qwen3-1.7B model from 8.5% to 32% on the GPQA scientific reasoning benchmark in under 10 hours on a single H100.
Paperclip, a new MIT-licensed Node.js platform, lets builders organize autonomous AI agents into company-style org charts with budgets, heartbeat scheduling, governance controls, and full audit logging. It supports any agent runtime and has crossed 30,000 GitHub stars since its March 2026 launch.
Moonshot AI open-sourced Kimi K2.6, a 1 trillion parameter MoE model that orchestrates up to 300 specialized sub-agents in parallel and runs autonomous coding sessions lasting 12+ hours. It scores 58.6 on SWE-Bench Pro, beating GPT-5.4 and Claude Opus 4.6, and ships with weights on Hugging Face under a Modified MIT License.
Alibaba shipped the Qwen 3.6 model family across April 20-22, including a proprietary Max-Preview variant that ranks first on six coding and agent benchmarks and an open-weight 27B dense model under Apache 2.0. The Max-Preview uses a mixture-of-experts architecture activating only 3 billion of 35 billion total parameters per inference, competing on cost efficiency against GPT-5.4 and Claude Opus 4.7.
DeepSeek launched preview versions of its V4 model family on April 24, featuring 1 million token context as default across both V4-Pro (1.6T total parameters, 49B active) and V4-Flash (284B total, 13B active). The open-source models are trained on Huawei Ascend chips and benchmark between GPT-5.2 and GPT-5.4 on reasoning tasks, with dedicated agent optimizations for Claude Code, OpenClaw, and OpenCode.
SUSE announced MCP integration across its infrastructure portfolio, enabling AI agents to autonomously monitor, troubleshoot, and patch enterprise Linux and Kubernetes environments through standardized tool calls. Partners include AWS, Fsas Technologies (Fujitsu), n8n, Revenium, and Stacklok.
Photon released Spectrum, an open-source TypeScript SDK that lets developers deploy AI agents to iMessage, WhatsApp, Telegram, Slack, and Discord with a single codebase. Developers write agent logic once and add platforms by changing one line. The SDK ships with sub-250ms message latency on Photon's edge network and has already processed over 400,000 messages in production through a matchmaking agent on iMessage.
Researchers at the University of Chicago launched Agent4Science, a Reddit-style social network where AI agents autonomously share, debate, and review scientific papers. Humans can observe but cannot participate. The platform has generated 40,000 comments from more than 150 agents across AI safety, deep learning, and related topics. It joins a growing wave of agent-exclusive platforms including Moltbook and EinsteinArena.
Chinese AI lab Moonshot AI shipped Kimi K2.6, an open-source model that coordinates up to 300 sub-agents running thousands of parallel steps for hours or days without human intervention. One internal team ran a K2.6 agent autonomously for five straight days handling monitoring and incident response. The release exposes a critical gap: most enterprise orchestration frameworks were built for agents that run for seconds, not days.
Cisco engineers detail a production multi-agent architecture that separates worker agents from leader agents, mirroring real engineering team structures. A pilot across 512 debug sessions by 70 users showed 93% reduction in time-to-root-cause and over 200 man-hours saved per month. Development workflows showed 65% faster execution.
SOLAI, formerly BIT Mining, is selling a $399 device that comes pre-loaded with OpenClaw and curated LLMs, designed to run autonomous agents 24/7 on local hardware with no coding required. It is the first consumer hardware product built specifically for self-hosted AI agents.
Cloudflare wrapped its first Agents Week with more than 30 product launches across five categories: compute primitives (Sandboxes GA, Artifacts versioned storage, Durable Object Facets), zero-trust security (Cloudflare Mesh, Managed OAuth, enterprise MCP governance), a full agent toolbox (Project Think SDK, voice agents, email service, AI Search, Agent Memory), prototype-to-production developer tools, and agentic web standards including an Agent Readiness score. The company is framing its Workers platform as 'Cloud 2.0,' purpose-built for a world where agents are the primary workload.
Tiiny Host released a Publish Agent Skill that lets coding agents like Claude Code, Cursor, and Gemini CLI deploy files to the live web from a single natural language instruction. The skill uses semantic intent matching rather than explicit commands, closing the gap between agent-generated output and live deployment.
Shopify's open-source AI Toolkit gives coding agents like Claude Code and Cursor live access to store operations, API schemas, and documentation through a Model Context Protocol server. Agents can update products, adjust inventory, and change pricing on live stores without touching the dashboard. The toolkit supports five AI coding tools and includes 16 skill files covering the full Shopify platform surface.
GitHub COO Kyle Daigle confirmed the platform now processes 275 million commits weekly, on pace for 14 billion in 2026. AI agents opened 17 million pull requests in March alone, up from 4 million six months prior. The surge has triggered visible outages as infrastructure built for human developers buckles under autonomous machine traffic.
Vercel released Open Agents, a forkable open-source platform that gives companies the web UI, agent runtime, sandboxed execution, and GitHub integration needed to build and run their own background coding agents. The project positions Vercel in a growing infrastructure race alongside Anthropic's Managed Agents and AWS Agent Registry.
NVIDIA dropped the full arXiv paper behind Nemotron 3 Super, its open-weight 120B model with 12B active parameters, 1M context window, and a hybrid Mamba-Transformer MoE backbone built specifically for multi-agent systems. The paper reveals how latent MoE, multi-token prediction, and native NVFP4 pretraining deliver over 5x throughput gains versus the previous Nemotron Super.
A production pilot at Cisco using LangChain's multi-agent framework compressed cross-team debugging from hours to under five minutes and cut development cycles by 65%. The gains came from coordination, not faster code generation.
Peter Steinberger, who built OpenClaw as a WhatsApp bot during a trip to Marrakesh, delivered a TED 2026 keynote on Saturday recounting how a personal experiment became one of the fastest-growing open-source projects in history. The talk marks a cultural milestone: AI agent infrastructure is now mainstream enough for TED's global stage.
FastGPT patched two NoSQL injection vulnerabilities on April 17 that, chained together, gave attackers unauthenticated root admin access and persistent account control over the 26,500-star AI agent building platform. The attack required nothing more than a crafted JSON payload with a MongoDB query operator.
Cloudflare and GoDaddy announced a strategic partnership on April 17 to build the identity and access control layer for AI agents on the open web. GoDaddy is integrating Cloudflare's AI Crawl Control into its hosting platform for its 21 million+ small business customers. Cloudflare launched isitagentready.com, a tool that scores any website on how well it supports AI agents, and a Cloudflare Radar dataset tracking agent standards adoption across the internet. Both companies are backing GoDaddy's Agent Name System (ANS), an open standard using DNS and PKI to give AI agents verifiable identities.
NanoCo's NanoClaw framework now integrates Vercel's Chat SDK and OneCLI's credential vault to ensure AI agents cannot execute high-stakes actions without explicit human approval delivered as native interactive cards inside 15 messaging apps. The architecture moves trust enforcement from the application layer to the infrastructure layer, so agents never see real credentials and cannot override the approval requirement regardless of prompt.
Microsoft's AI Toolkit for VS Code has been renamed to Foundry Toolkit and reached general availability. The GA release adds Agent Builder for low-code agent development, a catalog of 100+ models from OpenAI, Anthropic, Google, and local providers, configurable MCP tool approval, and one-click deployment to Microsoft Foundry Agent Service. An open-source agent-framework repository accompanies the release.
Chinese AI lab MiniMax simultaneously open-sourced M2 and shipped M2.7 today, a 230B-parameter mixture-of-experts model family designed specifically for agentic workflows. M2's API costs $0.30 per million input tokens and $1.20 per million output tokens, roughly 8-10% of Claude Sonnet 4.6's pricing, while running at approximately twice the speed. NVIDIA featured M2.7 on its Technical Blog, an unusual endorsement for an open-source release from a Chinese lab.
Databricks merged its AI Gateway into Unity Catalog, giving enterprises the same permissions, auditing, and policy controls over AI agent workflows that they already apply to data access. The release adds MCP governance with on-behalf-of user execution, guardrails powered by LLM judges, and per-request dollar cost tracking across all model providers.
Google's Android team released a new CLI, skills repository, and knowledge base designed specifically for AI agent workflows. Internal experiments showed 70% token reduction and 3x speed improvement over standard toolsets. The tools work with any agent, not just Gemini.
OWASP's GenAI Security Project released its Q1 2026 exploit round-up, the first comprehensive AI agent exploit taxonomy of the year. The report covers eight incidents from January through April 11, including the Mexican government breach via Claude-assisted attack workflows, OpenClaw inbox deletion, Meta's internal AI agent data leak, and active exploitation of Flowise CVE-2025-59528. Each incident is mapped to both the OWASP Top 10 for LLM Applications 2025 and Top 10 for Agentic Applications 2026.
Cal.com is the first major commercial open source project to name AI explicitly as the reason for going closed source. CEO Bailey Pumfleet told ZDNet that AI tools like Claude Opus can now scour public codebases for vulnerabilities faster than companies can patch them, making open source 'like handing out the blueprint to a bank vault.' The company simultaneously launched Cal.diy, a stripped MIT-licensed community fork for hobbyists.
The LangChain team is preparing to release version 1.0 of its core Python package, the first stable release of the most widely used AI agent development framework. The restructure adds LangGraph as a dependency, re-exports core primitives at the top level, removes deprecated modules, and consolidates documentation. The team is actively soliciting developer feedback via the official LangChain Forum before the release goes live.
Two AI agent frameworks disclosed high-severity remote code execution vulnerabilities on April 15, both rooted in the same architectural failure: trusting user-supplied input in MCP server configuration fields. CVE-2026-30617 affects LangChain-ChatChat 0.3.1 and CVE-2026-30624 affects Agent Zero 0.9.8, each carrying a CVSS score of 8.6 with no authentication required.
The world's dominant 3D design and manufacturing software platform just made its entire product line accessible to external AI agents. Autodesk's new Fusion MCP lets agents take action inside CAD workflows, not just read data, while Autodesk Assistant goes live across Fusion, Inventor, Moldflow, Vault, and Revit 2027 with natural language access to engineering tools.
Cloudflare's Project Think previews the next generation of its Agents SDK with durable execution, sub-agents, persistent sessions, and real-time voice interactions in roughly 30 lines of server-side code. The batteries-included base class wires all primitives together, positioning Cloudflare's agent infrastructure for long-running, voice-capable production deployments.
Cloudflare's Agents Week peaked on April 15 with four new product launches: an Agents SDK v2 preview for long-running multi-step tasks, Agent Lee (an in-dashboard agent replacing manual navigation), Cloudflare Mesh for secure private networking between agents and internal infrastructure, and a unified cf CLI covering all 3,000 Cloudflare API operations. Combined with Sandboxes GA and Dynamic Workers from earlier in the week, Cloudflare has assembled the most complete agent infrastructure stack from any non-AI-lab vendor this quarter.
GitHub Security Lab released Season 4 of its free, open-source Secure Code Game on April 14, themed 'Hack the AI Agent.' Five progressive challenges teach developers to exploit and defend against real-world agentic AI attack surfaces, from sandbox escapes to multi-agent trust chain exploitation. The game was inspired directly by OpenClaw's capabilities and the security risks they introduce.
Honor became the first major hardware OEM to ship a Lobster-architecture AI agent at the factory level on April 13. YOYO Claw comes pre-installed on MagicBook laptops with 23 specialist sub-agents, a dedicated security module, and a claimed 50% token reduction compared to OpenClaw.
Google Research published two multi-agent frameworks targeting academic publishing bottlenecks. PaperVizAgent orchestrates five specialized agents to generate publication-ready figures from manuscript text, scoring 60.2 on a calibrated evaluation scale where human performance sits at 50.0. ScholarPeer automates peer review using adversarial literature search and multi-aspect verification, producing reviews that are more critical and literature-grounded than existing automated approaches.
Microsoft released the Agent Governance Toolkit as an open-source project under MIT license, a nine-package monorepo that applies operating system security patterns to autonomous AI agents. The toolkit intercepts agent actions at runtime with deterministic policy enforcement, assigns cryptographic identities using decentralized identifiers, isolates execution through CPU-inspired privilege rings, and adapts SRE practices like error budgets and circuit breakers to agent workloads. It covers all 10 OWASP Agentic AI risks and ships with 20+ framework adapters for LangChain, AutoGen, Google ADK, and others.
The OpenClaw Hackathon 2026 ran April 11-13 and produced projects spanning autonomous content repurposing, live streaming agents, AI-managed social media profiles, and ambient voice assistants. The recurring theme across projects was not capability but infrastructure: observability, persistent memory, and access control are what builders need to move agents from demos to production deployments.
Google shipped ADK Python v1.30.0 overnight with Gemma 4 model support, Auth Provider integration for agent registries, artifact streaming in A2A protocol events, and a credential leakage fix. The most active open-source multi-agent framework in 2026 now lets agents pass live artifacts in agent-to-agent communications and manage credentials across distributed agent ecosystems.
Researchers at UC Berkeley's Center for Responsible Decentralized Intelligence built an automated agent that exploits eight of the most widely cited AI benchmarks to achieve near-perfect scores. No reasoning. No LLM calls. Just pytest hooks, binary trojans, config leakage, and sandbox escapes. The findings mean any published agent benchmark score is suspect without independent verification.
Two related Anthropic developer crises exploded on Hacker News on April 12 with over 1,100 combined comments. Developers documented that Claude Code v2.1.100+ silently injects approximately 20,000 invisible server-side tokens per request, exhausting Pro Max quotas in as little as 90 minutes of moderate use. Separately, forensic analysis of 119,866 API calls revealed Anthropic silently downgraded prompt cache TTL from one hour to five minutes on March 6, inflating cache creation costs by 17% with no public announcement.
Google Colab's new Learn Mode turns Gemini into a step-by-step coding tutor that understands full notebook context. Separately, Gemini CLI v0.33.0 shipped authenticated agent-to-agent communication, addressing a key gap in secure multi-agent orchestration.
NousResearch released Hermes Agent, an open-source autonomous agent that creates skills from experience, improves them during use, and builds a persistent model of each user across sessions. It runs on any model provider, deploys to serverless infrastructure, and connects to Telegram, Discord, Slack, WhatsApp, and Signal through a single gateway process.
OpenClaw 2026.4.10 promotes Codex from a compatibility-layer bolt-on to a native provider with its own auth, threads, and compaction. A new Active Memory plugin proactively surfaces relevant context before agent replies. The release also adds local MLX speech on macOS and tightens security across browser, sandbox, tool, and startup boundaries.
GitHub released the Copilot SDK in public preview on April 11, a multi-language SDK spanning Python, TypeScript, Go, .NET, and Java that exposes the same agent runtime powering Copilot CLI. Developers define agent behavior and the SDK handles planning, tool invocation, and file edits autonomously. MCP integration, BYOK authentication, and community ports for Rust, Clojure, and C++ are included.
Microsoft released Agent Framework on April 11, an open-source framework for building, orchestrating, and deploying AI agents in both Python and .NET. The framework includes graph-based workflows with streaming, checkpointing, human-in-the-loop controls, and time-travel debugging. Migration guides from AutoGen and Semantic Kernel position it as Microsoft's unified agent development platform.
Multica treats AI agents as first-class team members on sprint boards. Create an issue, assign it to an OpenClaw agent, and watch it execute autonomously. No vendor lock-in, no cloud required.
LM Studio has acquired Locally AI, an app for running AI models locally on iPhone, iPad, and Mac. Creator Adrien Grondin joins LM Studio to lead native cross-device AI experiences. The company says it plans to ship 'new ways to use your models and agents seamlessly across your own devices.'
Z.AI's GLM-5.1 scored 58.4 on SWE-Bench Pro, beating Claude Opus 4.6, GPT-5.4, and Gemini 3.1 Pro. The 754-billion parameter Mixture-of-Experts model, released under MIT license, can sustain productive autonomous coding for 8+ hours and 1,700 consecutive steps. It is the first open-weight model to top commercial LLMs on agentic software engineering benchmarks.
Anthropic's latest Claude Code release adds a Focus view toggle that reduces terminal noise during agent runs, tightens Bash tool permission checks around environment variables and network redirects, fixes an MCP connection memory leak accumulating 50MB per hour, and resolves multiple resume and transcript reliability issues. The update signals a push toward production readiness for enterprise coding agent deployments.
In the last week of March, LangChain disclosed three high-severity CVEs affecting 60 million weekly downloads, Langflow was exploited within 20 hours of disclosure, and a threat group hijacked LiteLLM's PyPI publishing pipeline to distribute credential-stealing malware. A new analysis argues these aren't isolated incidents. They're symptoms of an infrastructure class that grew faster than its security posture.
CrewAI v1.14 eliminates its LangChain dependency entirely, rebuilt from scratch with its own orchestration logic. The framework now ships standalone with 40,000+ GitHub stars, 60+ built-in tools, and native support for any LLM provider. The move positions CrewAI as a direct competitor to LangChain rather than an extension of it.
LangChain's Deep Agents Deploy entered beta on April 9, offering model-agnostic agent deployment with full memory ownership on self-hosted infrastructure. The framework directly challenges Anthropic's Claude Managed Agents, launched just one day earlier, by letting teams keep agent data in their own databases rather than a proprietary system.
Tether's new QVAC SDK lets developers build AI agents that run entirely on-device across iOS, Android, Windows, macOS, and Linux, with built-in peer-to-peer primitives for decentralized inference. The crypto company's bet: centralized cloud AI won't scale to trillions of agents.
Threat actors are actively exploiting a maximum-severity remote code execution flaw in Flowise, the open-source platform for building AI agents. CVE-2025-59528 scores a perfect 10.0 on CVSS and lets attackers execute arbitrary JavaScript on the server through the CustomMCP node. With 12,000+ instances exposed on the internet, this is the third Flowise vulnerability to see in-the-wild exploitation.
PraisonAI, the open-source multi-agent framework, disclosed and patched two critical remote code execution vulnerabilities within 24 hours. CVE-2026-39888 scores 9.9 on CVSS and allows sandbox escape through exception frame traversal. CVE-2026-39890 scores 9.8 and enables arbitrary code execution via malicious YAML agent definitions. Both are fixed in the latest releases.
OpenClaw's latest release adds a grounded REM backfill system that replays old daily notes into durable memory, patches SSRF and node exec injection vulnerabilities, introduces character-vibes QA evaluations, and overhauls Android device pairing from scratch.
Cursor's Bugbot code review agent can now learn from developer reactions and comments on pull requests, automatically promoting effective review rules and disabling stale ones. The feature has generated over 44,000 learned rules across 110,000 repositories, pushing Bugbot's resolution rate to 78%, fifteen points ahead of its nearest competitor.
Atlassian is embedding third-party AI agents directly into Confluence pages using Model Context Protocol. Lovable, Replit, and Gamma are the first three, turning product specs into prototypes, technical docs into starter apps, and notes into presentations without leaving the wiki.
A wave of developer comparison guides positions Eigent, a CAMEL-AI-based multi-agent desktop platform, as the first open-source alternative to OpenClaw generating sustained editorial attention. The two platforms represent fundamentally different bets on what an AI agent should be: chat-native personal assistant versus visual multi-agent workforce.
Microsoft shipped Agent Framework v1.0 on April 3, combining AutoGen and Semantic Kernel into a unified SDK for building multi-agent systems in .NET and Python. The production-stable release supports seven model providers and ships with A2A and MCP interoperability. Developer reaction, per Forbes, is skeptical: rivals like LangChain and OpenClaw are winning on simplicity while Microsoft's combined stack still feels like two frameworks wearing a trench coat.
Nous Research has published a comprehensive migration guide for moving from OpenClaw to Hermes Agent, its MIT-licensed autonomous AI framework that launched in February 2026 and has collected 22,000 GitHub stars. The guide includes a one-command migration tool. Hermes Agent's core differentiator is a closed learning loop: the agent writes reusable skill files after completing tasks, stores outcomes in persistent memory, and improves without manual configuration.
LangChain's official engineering blog defines three layers where AI agents can learn — model, harness, and context — and names OpenClaw's SOUL.md as the working example of agent-level context learning. The framework giant is telling developers the next generation of agents rewrites its own configuration over time.
Chinese users built custom AI agents called 'lobsters' on top of OpenClaw's open-source stack because Claude and ChatGPT are blocked in China. One IT engineer processes 200 TikTok Shop listings in two minutes instead of twelve a day. Government subsidies in Shenzhen, Wuxi, and other cities incentivize adoption — then Beijing's cybersecurity authorities warned of security risks and started pulling it back. The deeper story is what happens when open-source agent infrastructure meets a market locked out of Western frontier models.
Google released ADK for Java 1.0.0 on March 30 and ADK for Go 1.0 on March 31, making it the only major agent framework with four language SDKs. A new Morph comparison of eight agent frameworks surfaces a widening split: provider-native SDKs from Anthropic, OpenAI, and Google trade model flexibility for deeper integration, while independent frameworks like LangGraph and CrewAI offer cross-provider portability at the cost of native features. Meanwhile, IBM's ACP has merged into Google's A2A under the Linux Foundation, consolidating the agent-to-agent protocol layer.
Andrej Karpathy published an approach to AI agent memory on April 3 that ditches vector databases and RAG pipelines in favor of a structured Markdown wiki that the LLM actively compiles, links, and maintains. For teams building agents that need persistent project memory across sessions, the architecture addresses the core pain: context-limit resets that wipe everything the agent has learned.
Three days after Anthropic banned OpenClaw users from Claude subscriptions and required them to pay extra for API access, Google launched Gemma 4 under Apache 2.0 — a capable open-weight model that runs locally via Ollama. LushBinary published a guide pairing the two this weekend. The combination gives builders a zero-cost, privacy-first alternative with no API bills and no data leaving their machine.
Anthropic published a new engineering paper detailing a three-agent architecture that separates planning, code generation, and evaluation into distinct agents with context resets between sessions. The design targets multi-hour autonomous coding runs where single-agent approaches consistently break down.
Zscaler ThreatLabz found a malicious GitHub repository posing as leaked Claude Code source code that delivers Vidar v18.7 and GhostSocks malware. The campaign mirrors a similar attack using fake OpenClaw packages that Huntress flagged in March, with the same two payloads targeting developers searching for leaked or open-source AI agent tools.
Anthropic issued a DMCA takedown notice targeting repositories hosting its leaked Claude Code source code, but the request swept up 8,100 repositories including legitimate forks of Anthropic's own public Claude Code repo. The company retracted the bulk of the notices within hours, limiting enforcement to one repository and 96 forks. The incident compounds a difficult stretch for a company reportedly planning an IPO, and carries an uncomfortable irony: Anthropic has faced multiple copyright lawsuits over its own use of copyrighted training data, including a $1.5 billion damages order in September 2025.
GitHub now scans code changes for exposed secrets inside AI coding agent sessions through its MCP Server, treating agentic coding as a distinct security surface. The company also added 37 new secret detectors across 22 providers in March, with push protection now active by default for 39 token types. The move comes as GitGuardian reports 24,008 unique secrets exposed in MCP-related configuration files across public GitHub.
An open-source project called Claw Code launched today as an AI coding agent framework built in Python and Rust. The project claims 72,000 GitHub stars within its first days, though the actual count is difficult to verify independently. Built as a clean-room reimplementation of architectural patterns from the leaked Claude Code source, Claw Code positions itself as an auditable alternative to proprietary AI coding tools like Copilot and Cursor.
A packaging error in Claude Code version 2.1.88 shipped a .map source file that exposed nearly 2,000 TypeScript files and 512,000 lines of readable source code for Anthropic's CLI coding agent. The leaked code reveals anti-distillation mechanisms that inject fake tool definitions to poison competitor training data, an 'undercover mode' that strips Anthropic identifiers from AI-authored open-source commits, frustration-detection regex patterns, native client attestation via Zig-level HTTP hash injection, and a background agent mode called KAIROS. Anthropic confirmed the incident to CNBC, calling it a 'release packaging issue caused by human error.' The code was mirrored to GitHub before the npm package was pulled and has already surpassed 84,000 stars.
Microsoft's Azure Developer CLI (azd) shipped seven releases in March 2026, headlined by a new AI agent extension that lets developers run, invoke, and monitor agents locally before deploying to Microsoft Foundry. The update also includes GitHub Copilot integration for AI-assisted project scaffolding and error troubleshooting, Container App Jobs deployment, and local preflight validation. For agent builders on Azure, this closes the gap between writing agent code and testing it against live infrastructure.
TechRadar Pro published a roundup of 10 creative OpenClaw projects built by the community, ranging from multi-agent software development pipelines and overnight research systems to a WHOOP wearable tracker on a Raspberry Pi and Moltbook, a social network for AI agents that Meta acquired in March 2026 for its agent-to-agent communication infrastructure.
Ollama has released Pi — the minimal coding agent built on approximately 4,000 lines of TypeScript that powers the OpenClaw framework — as a standalone, customizable tool. Developers can now spin up Pi directly from the Ollama CLI with zero configuration, choose from 2,000+ models across providers, and extend it with custom plugins, skills, and prompt templates. The default cloud model is Kimi K2.5, a 1-trillion-parameter MoE model priced at roughly 9x cheaper than Claude Opus 4.5.
Security researcher Niels Provos launched IronCurtain, an open-source project that isolates AI agents in virtual machines and enforces plain-English security policies before they can touch user accounts. The project, covered by Wired in February and now endorsed by Kaspersky as a potential blueprint, addresses the core tension in agent platforms like OpenClaw: the same broad access that makes agents useful also makes them dangerous.
Agent-Infra's AIO Sandbox launched this weekend as the 13th entrant in a market that barely existed a year ago. E2B has processed over 200 million sandbox sessions, and roughly half the Fortune 500 now runs agent workloads on isolated execution platforms. Cloudflare shipped Dynamic Workers that spin up isolated code execution 100x faster than containers. NVIDIA's OpenShell enforces system-level security policies that agents can't override. Fly.io's Sprites offer persistent VMs with sub-second checkpoint/restore. And a YC X26 startup called Microsandbox built credential isolation directly into the network layer. The question 'where should AI-generated code run?' has become a full-blown infrastructure category, and the market is already splitting into competing architectural philosophies that will shape how every production agent operates.
Nvidia released ProRL Agent, an open-source infrastructure that decouples reinforcement learning training from agent rollout by exposing rollout as a standalone API service. Integrated into NeMo Gym, the system achieved near-linear throughput scaling and boosted a Qwen3-8B model from 9.6% to 18.0% on SWE-Bench Verified — gains from infrastructure alone, not bigger models.
OpenAI has updated Codex with a searchable plugin directory, MCP server integrations, and shareable skills bundles — the same stack that Anthropic's Claude Code has offered since late 2025. All three major coding agents now speak MCP, and the competitive surface is shifting from model quality to ecosystem breadth.
OpenClawd, the managed OpenClaw hosting service, shipped verified skill screening, runtime sandboxing, and credential isolation on March 26 after Koi Security's audit of the ClawHub marketplace found 341 malicious skills out of 2,857 — approximately 12% of the entire catalog. A February update raised that count to 824 malicious skills across more than 10,700 listings. The update is the first platform-level security response from within the OpenClaw ecosystem, arriving in the same week as Cisco DefenseClaw and Sysdig's runtime enforcement announcements.
Web3 protocol Luffa announced today it has integrated OpenClaw as its AI agent layer, becoming the first platform to assign decentralized identities (DIDs) to AI agents. The integration gives OpenClaw agents verifiable on-chain identity, auditable behavior logs, and governable permission boundaries — a direct response to the 'permission black box' problem in current agent deployments.
A threat actor tracked as TroyDen is running a malware campaign through fake GitHub repositories that impersonate OpenClaw tooling. The repos use AI-generated READMEs, throwaway accounts to inflate stars and forks, and a split-payload LuaJIT trojan that evades sandbox analysis. Netskope Threat Labs has identified over 300 malicious packages connected to the same attacker infrastructure.
On March 24, a threat actor called TeamPCP pushed backdoored versions of LiteLLM to PyPI, embedding a three-stage credential stealer that harvested SSH keys, cloud tokens, and Kubernetes secrets from every environment where the package was installed. LiteLLM sits in the dependency chain of nearly every major AI agent framework, and Wiz estimates it is present in 36% of all cloud environments. The attack is part of a broader campaign that has already hit Trivy, Checkmarx, and multiple package registries, with TeamPCP now claiming collaboration with the extortion group LAPSUS$.
NousResearch released Hermes Agent, an MIT-licensed autonomous agent framework that creates skills from experience, maintains persistent memory across sessions, and runs on everything from a $5 VPS to a GPU cluster. It supports Telegram, Discord, Slack, WhatsApp, and Signal from a single gateway process, positioning it as a direct competitor to OpenClaw in the open-source agent space.
Cisco unveiled DefenseClaw at RSA Conference 2026, an open-source framework that scans AI agents for vulnerabilities, manages MCP server permissions, and quarantines compromised skills in two seconds. The tool builds on NVIDIA's OpenShell sandbox and was directly inspired by Cisco engineers running OpenClaw at home. GitHub availability is set for March 27.
SkyPilot engineers pointed Anthropic's Claude Code agent at Andrej Karpathy's open-source autoresearch framework and gave it access to 16 GPUs on a Kubernetes cluster backed by CoreWeave. Over eight hours, the agent autonomously submitted approximately 910 ML experiments, discovered that scaling model width outperformed every hyperparameter tweak combined, and taught itself to exploit performance differences between H100 and H200 GPUs without being told they existed.
WordPress.com opened its platform to third-party AI agents that can autonomously write, edit, publish, and manage posts and pages — a move that could reshape content creation across the 40% of the web that runs on WordPress.
OpenAI will acquire Astral, the company behind Python's uv package manager and ruff linter. The Astral team will join Codex, which now has over 2 million weekly active users. This is OpenAI's second acquisition in March, following the OpenClaw deal, and both are pending regulatory approval.
OpenClaw creator Peter Steinberger publicly accused Tencent of scraping all skills from ClawHub and importing them into its QClaw platform without attribution or financial support. Within days, Tencent Cloud and Tencent AI became official OpenClaw sponsors — but the episode exposes a structural problem in open-source AI economics.
Y Combinator CEO Garry Tan released his personal Claude Code configuration — a collection of AI 'skills' that simulate an engineering team — and it went viral. 20,000 GitHub stars, 2,200 forks, a Product Hunt trending spot, and a developer backlash that exposed a genuine fault line in how software gets built in 2026.
Nvidia launched two new Nemotron 3 open models at GTC 2026: a 4-billion-parameter Nano model for consumer GPUs and a 120-billion-parameter Super model for professional workstations. Combined with optimizations for Qwen 3.5 and Mistral Small 4, Nvidia is building the open-model layer for local AI agents.
Open-source model unifies reasoning, multimodal, and agentic coding in a single Apache 2.0 model with 256K context window. Partners with NVIDIA for frontier model co-development.
OpenClaw has 150,000+ GitHub stars, triggered a CNCERT security advisory, and launched an enterprise arms race at Alibaba. Its Austrian creator remains largely invisible in the coverage.
Microsoft's new Agent Framework consolidates its enterprise AI tooling, but original AutoGen contributors have forked the project as AG2, raising familiar questions about Big Tech stewardship of open-source AI.
Amazon gives OpenClaw its first official cloud blueprint — the same week security researchers flag one in five ClawHub marketplace skills as malware. Two realities of the same platform, colliding in real time.
A new OpenClaw framework update makes Agent Communication Protocol bindings persistent across restarts, eliminating a major pain point for multi-agent production workflows.