Goals

July 16, 2026 ยท View on GitHub

Product outcome

AgentOps makes one coding-agent experiment independently inspectable without taking over the consumer's engineering system.

The canonical outcome is:

explicit behavior
  -> bounded implementation experiment
  -> exact content identity
  -> fresh independent judgment
  -> durable PASS | FAIL | NOT_PROVEN

Fitness properties

  1. Behavior before activity. The caller-owned intent states the active behavior, acceptance examples where useful, non-goals, evidence, and bounded write scope before implementation begins.
  2. One experiment. Implement performs one RED -> GREEN -> refactor cycle and reports facts without retry or delivery authority.
  3. Fresh judgment. PASS requires explicit, distinct author and validator context IDs plus a freshness attestation.
  4. Exact subject. Content identity covers files, symlinks, deletions, executable bits, roots, exclusions, and canonical digests without requiring Git or ao.
  5. Honest uncertainty. Mutation, incomplete changed-path coverage, missing identity, or missing proof returns NOT_PROVEN. Proven scope or acceptance failure returns FAIL.
  6. Sovereign proof. Validate atomically writes content-addressed JSON to caller-controlled storage. Provenance is optional audit.
  7. Stop boundary. RPI dispatches Plan, Implement, and Validate at most once, reports the result, and emits no next action.
  8. Open ecosystem. Callers keep their trackers, Git, PRs, CI, cloud agents, merge queues, rollback, and release systems.

Structural constraints

  • Core hard dependencies are only rpi -> {plan, implement, validate}.
  • Learn and all strategy/factory/specialist skills are off-path or optional.
  • Core schemas contain no retry, budget, queue, claim, lease, admission, next-action, closure, release, or delivery state.
  • The pure manifest and verdict helpers make no Git, tracker, queue, network, release, or delivery call.
  • Deterministic ao gate check reports repository check success or failure only.
  • Semantic validation is a skill responsibility, not a CLI state machine.

Measured learning hypothesis

Collections of durable verdicts may reveal repeated defect classes that deserve better context, tests, or checks. Learn may propose evidence-backed candidates off-path. Promotion requires separate evaluation; no observation silently becomes policy or changes a prior verdict.