> ../patterns/skill_review_gate.md
§ 01 · The feature
Hermes Agent's defining capability is its learning loop: when it solves a hard problem, it writes a reusable skill document, and it refines existing skills as it keeps using them. Skills are searchable, shareable and follow the open agentskills.io format. For a personal agent, this is exactly what you want, it stops re-learning the same thing.
§ 02 · The production problem
In a production deployment, a skill is not a note. It is an instruction the agent will follow the next time a similar task appears, for every future user of that workflow. A skill created from one unusual case, or refined after a single success that happened for the wrong reason, changes how the agent behaves from then on. That is a change to production behavior with no author, no review and no test.
Most teams already have a rule for changes like that when they come from a person: code review and a passing test suite before it ships. The pattern here is simply to apply the same rule when the author is the agent.
§ 03 · The pattern
1. Separate learned from approved. Let the agent create and refine skills freely in a staging area, but only load skills from an approved set in the production workflow. Learning continues; deployment is gated.
2. Version every skill. Keep skills in version control so every change is a diff: what the skill said before, what it says now, and which session produced the change.
3. Review the diff, not the whole skill. A reviewer should read what changed and ask one question: would we accept this instruction if a new team member had written it?
4. Evaluate before promotion. Run the workflow's eval set with the new skill version. If results get worse, or merely change on cases they shouldn't touch, the skill stays in staging. See the eval set is the spec.
5. Make rollback one step. When a promoted skill misbehaves, reverting to the previous version should be a single, rehearsed action, not an investigation.
§ 04 · What this costs
It slows down how fast learned behavior reaches production, by design. In exchange, every change to what the agent does has a record, a reviewer and a test result. For a business workflow, that trade is almost always worth it; for a personal agent it usually isn't, which is why the default is different.
── Checklist ──
§ 05 · Checklist
[ ] Are learned skills kept separate from the skills production actually loads?
[ ] Is every skill under version control, with changes visible as diffs?
[ ] Does a person review skill changes before promotion?
[ ] Does the eval set run against a new skill version before it is approved?
[ ] Can you roll back a single skill in one step?
── End of pattern ──
◆ An agent that improves itself is valuable. An agent that changes its own production behavior without review is a system nobody fully owns.
ORBIRESEARCH