Skip to main content

Notes

  • The score comes from the task’s output (the agent self-scores), keeping gating synchronous with task completion, no engine change required.
  • For independent verification, attach an async LLM-judge scorer via <Task scorers={...}>. That score lands in _smithers_scorers for telemetry and is not consulted for gating.
  • currentScore = null | undefined means pending or missing; default onPending: "continue" runs another iteration so the next task’s output can populate the score.
  • <LoopUntilScored> thinly wraps <Loop>: the caller resolves the score from ctx, and the component only computes until and forwards it.