OpenClaw and Hermes Agree on What an Agent Is — They Disagree on What Controls It

A The New Stack comparative analysis (June 2026) that gives the spoke’s harness thread its cleanest frame: two open harnesses — openclaw and hermes-agent — share the same definition of an agent but bet on different control points. Fetched via the firecrawl skill (WebFetch returned only page chrome).

The shared definition (what both agree on)

The article’s crisp anatomy of a harness: the software that turns a model into a system that runs autonomously and unattended — a runtime (keeps the agent alive between tasks) + a gateway (messages in/out) + memory (persists across sessions), with tools, identity, skills, and policy/observability on top. The sharp line: a coding assistant like Claude Code or Codex “covers only part of this … loses most of its working context once that session ends,” whereas a harness “keeps the runtime, memory, and governance in place so the agent can run unattended.” (Echoes Jensen Huang’s “operating system for personal AI” and Nvidia’s “an agent is an LLM and a harness.”)

The disagreement — which part is the primary control point

• openclaw = gateway-first (breadth). Centered on a gateway connecting one agent to dozens of messaging channels (WhatsApp/Discord/Slack), with ClawHub, a marketplace of thousands of community skills. Steinberger’s project hit ~380k GitHub stars; it moved to an independent foundation with OpenAI as sponsor. Its gap was governance — which platform vendors supplied: Nvidia’s NemoClaw (OpenShell sandbox + external policy) and Microsoft’s Scout (microsoft-scout; Windows containers, Entra identity, Teams/Outlook/SharePoint).
• hermes-agent = memory-first (depth). Nous Research’s harness centers persistent layered memory: it learns a developer’s work, builds a profile, and refines its own skills after hard tasks — run on infra you own (VPS/home server/laptop). A hermes claw migrate command imports an OpenClaw user’s settings/memories/skills/keys in one step.

The thesis that lands in synthesis

• Breadth vs depth, and it’s an emphasis not an exclusivity — OpenClaw also has memory; Hermes also speaks ~20 channels. The real contest is the control layer, not the model.
• Memory is the durable lock-in. “An agent that has learned a year’s worth of a developer’s habits creates a higher switching cost than one that merely connects to many applications. Memory, more than channel reach, is becoming the durable form of lock-in.”
• The platform vendors want the runtime layer (it “will outlast any single foundation model”) — and it’s being built beneath both projects: NemoClaw already runs Hermes under the same controls as OpenClaw. Governance is becoming agent-agnostic infrastructure.
• Managed-vs-self-hosted, restated: OpenClaw-under-NemoClaw/Scout = vendor-governed convenience; Hermes = full control + operational responsibility.

Caveats

TNS analytical feature (T2): adoption metrics (stars, OpenRouter token volumes) “measure different kinds of adoption and rarely move together” (the article’s own caveat); both projects are “emerging platforms, not finished products.” Security is real: an audit flagged 341 malicious entries in OpenClaw’s skill marketplace — the supply-chain risk governed runtimes exist to close.

Related

openclaw · hermes-agent · agentic-coding-harness · microsoft-scout · claude-code-channels-vs-openclaw · agent-memory · agent-guardrails · agentskills-spec · agent-orchestration