agentsmyth

Appearance

agentsmyth in action ​

Illustrative walkthrough — the repo, the feature, and the specific findings below are fabricated but faithful, not a real transcript. Every artifact shape, gate, and rule shown is exactly how the workflow actually produces them.

One feature walking the whole road, from a one-line request to a reflected-on ship. The repo is a fictional payments API, the agent is Claude Code, the change is real-sized: add rate limiting to the public endpoints.

The request ​

You type one line. The adapter routes it to workflow/router.md, which classifies it against the rules.

text
> add rate limiting to the public API endpoints
router verdict: Complex — full loop mandatory, no shortcuts

The seven gates ​

  1. Think — the brief. Requirements captured with stable IDs. Q1 is a blocker, so the agent stops and asks. → briefs/rate-limiting-v1.md
  2. Plan — the decisions. Token-bucket limiter in middleware, backed by Redis. Verification criteria written before any code. → plans/rate-limiting-v1.md
  3. Build — the task. Changed files logged against the IDs they satisfy. Nothing outside scope is touched. → tasks/rate-limiting-v1.md
  4. Review — against the plan. Every active requirement confirmed; hardcoded Retry-After flagged. Verdict: pass-with-risk.
  5. Test — the evidence. The one phase that cannot be faked. Real command output recorded. No untested requirement reaches Ship.
  6. Ship — the risk signed. Deploy evidence recorded, rollback stated. The Review risk travels here visibly, not buried.
  7. Reflect — the loose ends. Learnings captured; the open follow-up refuses to leave without a named owner. It gets one.

The task artifact ​

As the agent implements, the artifact fills in real time. Its frontmatter is the machine-readable orchestration state:

yaml
---
manifest_ids: [R1, R2, RI1]
orchestration:
  phase: build
  status: in-progress
  next_phase: review
---
text
$ npm test -- rate-limit
PASS enforces 100 req/min per API key (R1)
PASS returns 429 over limit (R2)
PASS limit reload without restart (RI1)
verdict: ship

What you are left with ​

Seven files. The complete story of the change, in order, with the reasoning and the evidence attached.

text
workflow/artifacts/
briefs/rate-limiting-v1.md      R1 R2 RI1, A1 confirmed, Q1 answered
plans/rate-limiting-v1.md       token bucket, Redis, config-driven
tasks/rate-limiting-v1.md       2 files, mapped to IDs, evidence logged
reviews/rate-limiting-v1.md     pass-with-risk, Retry-After flagged
verify/rate-limiting-v1.md      three tests, real output
ship/rate-limiting-v1.md        deployed, risk signed, rollback stated
reflect/rate-limiting-v1.md     one learning, one owned follow-up

Send a different agent in next week to fix that Retry-After follow-up, and it will read this chain, know exactly what was decided and why, and pick up without asking you a single question. That is the difference between an answer and a record. That is loop engineering.