Hermes

Hermes is a self-improving AI Agent platform from Nous Research—personal-assistant oriented, with learning loops, messaging integrations, and Agent Skills support. We rate it hold with Moved Out: strong feature set and good fit with Ollama locally, but always-on personal agents that message external systems are inherently hard to secure; the effort to harden them often negates the convenience versus running a bounded agent in an IDE or scheduled automation you control.

Blurb

The self-improving AI agent built by Nous Research. The only agent with a built-in learning loop — it creates skills from experience, improves them during use, nudges itself to persist knowledge, and builds a deepening model of who you are across sessions.

Summary

Hermes targets the same problem space as OpenClaw—a always-available assistant across chat surfaces (Telegram, Slack, CLI, etc.) with memory, cron-style scheduling, and skill extensibility. It integrates Agent Client Protocol so editors like Obsidian can use the agent via plugins (e.g. obsidian-agent-client). On Apple Silicon, local runs via Ollama with models such as deepseek and qwen are practical.

Originally a credible competitor in the personal-agent wave; our position is that broad tool-and-shell access without tight policy boundaries is unsafe for most setups. Prefer IDE-bound agents (Claude Code, Cursor-class tools) or pipelines that emit reviewable scripts on a schedule—see ADR-0003-each-machine-runs-pipelines-independently for why we avoid shared always-on agent gateways across machines.

Details

• Strengths: skill learning loop, multi-channel gateway, Agent Skills portability, ACP for IDE clients, flexible model routing (cloud or Ollama).
• Risks: persistent credentials, messaging exfiltration, autonomous cron/subagent actions—treat as high-trust workload only with explicit sandboxing.
• When hold is OK: isolated experimentation, local-only models, single-user machine with clear data boundaries.
• When to avoid: production secrets on the same host, multi-tenant boxes, or “set and forget” automation without human review.
• Install: upstream documents quick install via install.sh on macOS/Linux; verify signing and update channel before production use.