Hermes Agent with OpenShell: skills and sandbox demo

Nous Research demonstrates Hermes Agent inside an OpenShell sandbox using the NemoClaw blueprint. The AI agent reviews a GitHub pull request, creates a replacement PR while preserving commit authorship, and saves the workflow as a reusable skill. A test fails during the demonstration; the presenter says the focus is workflow reuse rather than finding bugs.

A fresh session searches earlier sessions and retrieves the skill for a second PR. The presenter clarifies that this happens in Alice's sandbox, so the demo does not establish that Bob used Alice's skills. The team describes capped memory, a skill index that loads full instructions when needed, and a scheduled curator that consolidates the skill library. Here, self-improvement means retaining procedures and context across sessions.

OpenShell provides the policy boundary in this setup. The demo shows a placeholder instead of a GitHub token inside the sandbox and describes credential substitution outside it. A request to Brave's search endpoint is blocked, then connects after a policy update, though no Brave API key is configured.

The speakers call the enterprise combination a prototype and recommend testing permissions before production use. They also introduce Hermes Desktop as an Electron interface for personal use, distinct from OpenShell's security and tenant controls.