HeyRocky

Use Hermes Standing Goals Without Unbounded Continuations

A standing goal asks Hermes to judge each response against an active objective. If the objective is not satisfied, the runtime feeds a continuation prompt into the same session and keeps working. The loop ends when the goal is met, the continuation cap is reached, or the user pauses or clears it.

goals:
  max_turns: 20

The default cap of 20 continuation turns protects against fuzzy goals, unreachable outcomes, and judge false negatives. When the cap is reached, Hermes auto-pauses the goal and asks for an explicit resume.

Write a verifiable goal

A strong goal names:

  • the exact deliverable;
  • the allowed workspace or system;
  • completion evidence;
  • safety and authorization boundaries;
  • conditions that should pause rather than improvise.

“Improve everything” is a poor standing goal. “Repair the local test suite in this repository, run the documented checks, and stop before deployment” is bounded and verifiable.

Understand fail-open judging

A judge failure does not wedge the workflow; Hermes continues. That keeps transient evaluation errors from halting useful work, but it makes the turn cap essential. The goal judge is a progress mechanism, not a substitute for hard authorization, tool approvals, or external quotas.

Goal checklist

  • The outcome can be demonstrated with fresh evidence.
  • The workspace and non-goals are explicit.
  • Paid, destructive, or public actions retain separate authorization.
  • goals.max_turns is finite.
  • The operator knows how to pause, resume, and clear the goal.
  • A blocked prerequisite causes an honest stop.

Common pitfalls

  • Using a subjective goal with no acceptance test. The judge may continue even after useful work is complete.
  • Raising the cap instead of narrowing the goal. More turns amplify ambiguity.
  • Treating the judge as an approval system. It only assesses completion.
  • Forgetting session context. The active goal operates in one session and should use the correct working directory.

Verification

Set a harmless goal with a clear two-step outcome in a disposable workspace. Confirm Hermes continues after an intentionally incomplete first response, then stops when fresh evidence satisfies the objective. Test the cap with a deliberately impossible sandbox goal and verify it auto-pauses rather than running indefinitely.

Official references: Hermes configuration and standing goals.

Related pages