The race for the agent runtime isn't about models. It's about who controls the layer that keeps an agent alive, gives it memory, and decides what it can touch.

Two open projects defined that layer in 2026. OpenClaw, built around a broad gateway connecting agents to dozens of messaging channels, drew OpenAI, Nvidia, and Microsoft into its orbit. Hermes Agent, from Nous Research, built around persistent memory that learns a developer's codebase and refines itself over time — and overtook OpenClaw in OpenRouter's daily token rankings in May.

They agree on what an agent harness is. They disagree on which part matters most.

What actually changed

OpenClaw went enterprise via platform vendors. Nvidia wrapped it in NemoClaw at GTC in March, sandboxing each agent and enforcing policy from outside the agent's reach. Microsoft made it native to Windows execution containers at Build in June, shipping Scout — an enterprise agent with an Entra identity, wired into Teams, Outlook, and SharePoint. Breadth got distribution; the platform vendors added the controls.

Hermes built depth via memory. Released February 25 under MIT license, Hermes keeps a layered memory across sessions, develops new skills after hard tasks, and refines them with use. It builds a profile of the developer it works with — so each session starts with more context than the last. By late June, it sat at 22 trillion tokens on OpenRouter's app rankings, first by total usage.

Hermes also ships a migration command. hermes claw migrate imports an OpenClaw user's settings, memories, skills, and keys in a single step. That's not a feature — it's a land grab.

What this means

The analogy holds: this is managed cloud vs. self-managed infrastructure. OpenClaw is the managed path — platform-governed, vendor-controlled, increasingly integrated into enterprise tooling. Hermes is the self-hosted path — you own the infrastructure, you own the memory, you own the switching cost.

"Memory, more than channel reach, is becoming the durable form of lock-in."

That's the crux. An agent that's learned a year of a developer's habits, conventions, and decisions is far stickier than one that merely connects to many applications. NemoClaw already runs Hermes agents alongside OpenClaw agents — the governance layer is being built beneath both projects, not betting on one.

The security audit that flagged 341 malicious skills in ClawHub's marketplace and tens of thousands of exposed instances earlier this year tells you something too: distribution without governance is a liability. The platform vendors showed up precisely to fix that.

What to do

• Enterprise teams evaluating agents: Ask before either harness touches production — who can explain a change in agent behaviour between sessions, and who owns the policy engine and the agent's identity?
• Developers choosing a harness: Need channel breadth and vendor-governed guardrails? OpenClaw + NemoClaw or Scout is the path. Need long-lived context and model-agnosticism across hundreds of providers? Hermes is worth a proper look.
• Platform engineers: The runtime layer is where vendor lock-in is settling. hermes claw migrate already works — the projects are converging faster than the star counts suggest.
• Watching both: The next phase turns on ownership. Whichever project controls memory and governance at scale controls the enterprise agent market.

Source: OpenClaw and Hermes: Two Architectures Fighting for the Agent Control Layer — Janakiram MSV, The New Stack

✏️ Drafted with KewBot (AI), edited and approved by Drew.