Nous Research released Hermes Agent V0.12 on April 30, shipping an autonomous Curator agent that maintains the platform’s skill library without human oversight, according to the project’s official release notes on GitHub. The update spans 1,096 commits, 550 merged pull requests, and contributions from 213 community developers.
The headline feature is the Curator, a background agent that runs on a seven-day cron cycle. It reviews every saved skill in the agent’s library, grades each one for usefulness, merges duplicates, prunes dead skills, and writes a report of every change. Archived skills are classified as either consolidated or pruned. Defense-in-depth gates protect bundled and hub skills from accidental mutation.
The Self-Improvement Problem
Persistent AI agents accumulate skills over time. Every complex task completion can generate a new saved skill. Without cleanup, agent skill libraries grow bloated with outdated, overlapping, and conflicting entries, a problem that wastes tokens (and money) on every inference call that loads the skill context.
Hermes V0.12 addresses this with two mechanisms. The Curator handles periodic bulk maintenance. A separate self-improvement loop, substantially rewritten in this release, evaluates what to save or update after each turn using rubric-based grading rather than free-form judgment. The loop now prefers updating the skill the agent just loaded rather than creating a new one, and it properly inherits the parent’s live runtime credentials. Both systems are restricted to the memory and skills toolsets so they cannot sprawl into unrelated agent operations.
Platform Expansion
Beyond the Curator, V0.12 adds Google Meet integration through a plugin that joins calls, enables captions, generates transcripts, and can follow up after calls end. Spotify gets native tool support with seven tools covering playback, search, queue management, and playlist operations.
Microsoft Teams arrives as the 19th messaging platform through a new pluggable gateway architecture, the first platform shipped as a drop-in adapter rather than a core integration. Tencent Yuanbao is the 18th. Four new inference providers (GMI Cloud, Azure AI Foundry, MiniMax, and Tencent Tokenhub) join, alongside first-class LM Studio support for local model execution.
Performance and Tooling
Cold start time for the terminal UI dropped by approximately 57%, according to the release notes. The project also added a remote model catalog manifest so that new models on OpenRouter and Nous Portal appear automatically without requiring a version update.
ComfyUI v5 and TouchDesigner-MCP integrations moved from optional to bundled-by-default, positioning Hermes as a creative production tool alongside its productivity and messaging capabilities.
Competitive Context
Hermes Agent operates in the same space as OpenClaw: persistent personal agents that connect to messaging platforms and grow more capable over time. Both are open-source. The Curator release positions Hermes around a specific thesis: that long-running agents need autonomous self-maintenance to remain useful, not just human-driven configuration. Whether that seven-day cleanup cycle actually reduces skill rot at scale is an open question that will depend on how aggressively users customize their agents.