Agentic AI
OpenClaw and Hermes chart different paths to agent control—gateway vs. memory
Two leading agent harness projects agree on the basic architecture for autonomous AI but diverge on whether to prioritize message channels or persistent learning.
The New Stack
An agent harness—the software layer that keeps AI models running autonomously between tasks—is shaping up as the real battleground in enterprise AI, not model selection itself. Two open-source projects are competing to define it, starting from opposing assumptions.
OpenClaw, which counts OpenAI, Nvidia, and Microsoft among its backers, treats the gateway as primary: a single connection point that lets one agent answer across WhatsApp, Discord, Slack, and dozens of other channels. Hermes Agent, from Nous Research, starts with memory instead—allowing an agent to build persistent context and refine skills over weeks of work with a developer.
Read the full analysis: OpenClaw and Hermes agree on what an agent is. They disagree on what controls it
The anatomy of an agent harness#
Both projects agree on the core components: a runtime that keeps the agent alive, a gateway for messages, persistent memory across sessions, policy controls, and observability. The difference is which lever they pull first.
OpenClaw's breadth gives it reach—nearly 380,000 GitHub stars and native integration into Windows and Microsoft's Scout enterprise agent. Hermes' depth drives usage: it surpassed OpenClaw in OpenRouter's token rankings by May, with developers favoring its ability to carry context and develop skills over time.
The real prize: governance and identity#
For platform vendors, the architecture choice matters less than the layer beneath it. Nvidia wrapped OpenClaw in NemoClaw, adding sandboxing and policy enforcement. Microsoft made OpenClaw native to Windows containers and shipped Scout with Entra identity. Both effectively placed governance, identity, and observability in the platform vendor's control—whether the agent above runs OpenClaw or Hermes.
That shift has real enterprise value. Early OpenClaw deployments risked broad access to files and systems. A governed runtime lets security teams scope permissions tightly and audit agent behavior.
The switching cost story#
The long-term competition will hinge on lock-in. An agent that has learned a year of a developer's habits and conventions creates stickiness that channel reach alone cannot match. Memory, not breadth, is becoming the durable form of dependency. That's why OpenRouter's token rankings now show Hermes leading OpenClaw—developers are choosing depth.
