OpenClaw and Hermes Are Building the Agent Runtime Layer From Opposite Ends

OpenClaw and Hermes agree on what an agent harness does: it turns a language model into a system that can run continuously, remember what it learns, and call tools to act. They disagree on which part of that system should be the primary control point. The New Stack published a detailed architectural comparison on June 23 that maps the divergence and its implications for enterprise adoption.

Two Projects, Two Control Models

OpenClaw, the open-source harness created by Peter Steinberger and now backed by an independent foundation with OpenAI as a sponsor, is built around the gateway. One agent can answer on WhatsApp, Discord, Slack, and dozens of other channels from a single runtime. ClawHub, its public skills marketplace, holds thousands of community skills that extend what the agent can do. The repository has reached approximately 380,000 GitHub stars by late June, according to The New Stack’s reporting.

Hermes Agent, released by Nous Research on February 25 under an MIT license, is built around persistent memory. The agent keeps a layered memory across sessions, develops new skills after completing hard tasks, and refines those skills over time. It also builds a profile of the developer it works for, so each session starts with more context than the last. Hermes surpassed 100,000 GitHub stars by mid-May and reached roughly 160,000 by late in the month. On OpenRouter’s daily token rankings, Hermes overtook OpenClaw on May 10, with 224 billion tokens processed that day, and by late June had placed first by total tokens past 22 trillion.

The distinction is one of emphasis rather than exclusivity. OpenClaw includes memory and skills. Hermes speaks across twenty-plus channels. But the architectural starting point shapes how each project handles governance, portability, and lock-in.

Platform Vendors Are Building Beneath Both

The more significant development is how platform vendors are positioning relative to both projects. At Microsoft Build this month, CEO Satya Nadella showed OpenClaw running natively on Windows inside execution containers, with Scout, Microsoft’s always-on enterprise agent, built on top of it. Scout has its own Entra identity and connections into Teams, Outlook, and SharePoint.

At GTC in March, Nvidia wrapped OpenClaw in NemoClaw, an OpenShell runtime that sandboxes each agent and enforces policy from outside the agent’s reach. Nvidia’s NemoClaw blueprints, according to The New Stack, already run Hermes agents under OpenShell as readily as they run OpenClaw. The governance layer is being built to sit beneath multiple agent projects rather than to pick one.

For enterprises, the practical impact is that a security team can now scope which folders an agent reads and which stay hidden, rather than granting the broad access that made early OpenClaw deployments risky. Platform teams can offer a single governed agent that staff access from existing tools.

Memory Creates Lock-in, Gateways Create Distribution

The New Stack’s analysis surfaces a subtler competitive dynamic. An agent that has learned a year’s worth of a developer’s habits creates a higher switching cost than one that connects to many applications. Memory, more than channel reach, is becoming the durable form of lock-in.

Nous Research has made portability part of its pitch, shipping a hermes claw migrate command that imports an OpenClaw user’s settings, memories, skills, and keys in one step. That command exists because Nous understands the lock-in dynamic and is trying to lower the barrier to switching while the market is still forming.

The tradeoff for Hermes users is operational. The team that runs Hermes also has to secure and maintain the infrastructure it lives on. OpenClaw’s gateway model distributes that burden across the platform vendors who have adopted it.

Security Gaps Remain in Both Ecosystems

The article notes that audits of OpenClaw’s skill marketplace flagged 341 malicious entries among the skills scanned, and security firms reported tens of thousands of exposed instances earlier this year. Those are the gaps that governed runtimes like NemoClaw and Microsoft’s execution containers are designed to close.

For enterprise buyers evaluating either project, The New Stack raises two questions worth answering before deployment. First, accountability: when an agent can rewrite its own memory and skills between sessions, as Hermes does, someone needs to explain a change in behavior and point to where that change is logged. Second, ownership: when governance and identity come from a platform vendor, the policy engine and the identity belong to that vendor, not to the team running the agent.

The Runtime Layer Outlasts the Model

The agent market is moving past model selection into the runtime, governance, and memory layers. OpenClaw demonstrated that a broad gateway and a large skills ecosystem can attract developers and draw in OpenAI, Nvidia, and Microsoft. Hermes demonstrated that persistent memory and self-improving skills can drive heavy daily usage without the same platform backing.

Whether breadth and depth remain in separate projects is already uncertain. NemoClaw runs both under a single set of controls, and Hermes can import an OpenClaw setup. Nvidia and Microsoft are competing to place governance, identity, and observability around whichever agent a customer chooses, because the runtime layer will outlast any single foundation model.