Hermes /learn — capturing workflows as skills without hand-writing SKILL.md

A MarkTechPost write-up (2026-06-24) of Nous Research’s new /learn command for Hermes Agent: point the live agent at “a document page, a local SDK, a past conversation, or pasted notes” and it authors a standards-compliant SKILL.md on your behalf — no hand-writing. A concrete mechanism for the self-improvement thesis, and a sharpening of the agentskills-spec skill model.

How it works

• Sources: local dirs (code/API docs/PDFs), web URLs, conversation history, or pasted text. Invocations like /learn the REST client in ~/projects/acme-sdk, focus on auth + pagination, /learn https://docs.example.com/api/quickstart, /learn how I just deployed the staging server.
• Mechanism — no separate ingestion pipeline: it reuses the agent’s own tools (read_file, search_files, web_extract) to gather material, then writes the SKILL.md.
• Storage/invocation: skills live in ~/.hermes/skills/, become slash commands automatically, and follow the agentskills.io format with three-tier progressive disclosure — Level 0 index (~3k tokens) → Level 1 full skill → Level 2 reference files.
• Approval gate: new skills pass through a skill_manage tool; with write_approval on they land in ~/.hermes/pending/skills/ for review — but the default is off, so the agent writes skills independently.

Why it matters — a skill-authoring autonomy spectrum

The piece’s cleanest contribution is naming four creation methods along an autonomy axis, which maps the self-improving-agents thread precisely:

1. Hand-write (you, manual) — full control.
2. /learn (live agent, you point it at material) — assisted conversion.
3. skill_manage auto (the agent itself, after a completed 5+ tool-call workflow) — unprompted procedural memory: the closed learning loop captures its own success as a reusable skill.
4. Hub install (third-party, registry/GitHub) — community.

So Hermes spans manual → assisted → autonomous → community skill creation in one system. The approval gate is a concrete answer to the drift/quality risk — the same gate self-authored changes discipline as zouroboros‘s three-model consensus vote, here a human-review queue (though defaulted off). And by emitting a standards-compliant SKILL.md regardless of method, it keeps the portable-skill format as the common output even as the author moves from human to agent.

Caveats

Trade-news relay of a vendor feature (T3): mechanism details are Nous’s; no evidence on quality of auto-authored skills (the open self-improvement question — does agent-written skill capture stay net-positive, especially with write_approval off by default?).

Related

hermes-agent · self-improving-agents · agentskills-spec · agent-skills · zouroboros · agent-memory · compound-engineering