AI Agent field notes

The agreement ladder

The tested artifact, verbatim, from the design project where the incidents in settled is a human word happened. The owner wrote it, revised it after adversarial review, runs it in that project’s instruction file, and reports it holding. Read it in that note’s terms: the ladder is vigilance support, not prevention — it makes rounding-up cheap to catch (rungs are greppable, quotes are checkable) and the catching stays the owner’s job.

Why it exists: when reviews or critics produce findings in bulk, each handoff quietly upgrades the level of agreement — a reviewer’s claim becomes a problem “we have,” an acknowledged problem becomes a fix “we chose,” a liked direction becomes a spec “we agreed,” a fresh draft becomes a settled artifact. No step lies; each rounds up half a rung, and after three steps the owner faces “decided” work he never decided.

Every task and decision sits on exactly one rung or terminal state, named in its status:

  • L0 claimed — a reviewer, critic, or the assistant asserts it; the owner has not spoken. Items are born here.
  • L1 issue confirmed — the owner agrees the problem is real. Nothing about the fix is agreed.
  • L2 direction agreed — the owner picked the shape of the fix. The details are not written.
  • L3 spec drafted — a written spec exists and is under review. It is a draft, whoever wrote it.
  • L4 ready — the owner signed the spec. Only now may it be built or folded into settled design docs.
  • Terminal: rejected (the owner ruled the claim invalid) or done. Both leave the ladder and move to the changelog with a closing note.

The rules:

  1. A critic’s finding is a claim: it is born at L0. The owner’s verdict is collected finding by finding, in conversation, before anything else happens to it.
  2. A rung changes only on the owner’s words, and the status QUOTES them. A quote is checkable; a date alone is not evidence and never suffices. If the words support L1, the status says L1 — even when the fix seems obvious.
  3. One utterance moves an item at most one rung, and never past the rung its words support. Accepting a problem is not accepting a fix; accepting a direction is not accepting a plan.
  4. Artifacts are born DRAFT. Any catalog, diagram, spec, or prompt the assistant produces carries draft status until the owner has walked through it.
  5. The decision log records only decisions the owner made. Recording an assumption, a reviewer’s conclusion, or the assistant’s own inference as “accepted” is a defect; withdraw it when found.
  6. Inference is not a ruling. “The thing you approved requires X” may be true — X still waits for the owner’s word, with the inference written as a note beside it.
  7. A verdict on unreadable text is not a verdict. Findings put to the owner must be readable: the concrete story, what exists today, the options, their costs. Compressed labels manufacture fake agreement.
  8. Who types the rung: the assistant may type L0 through L3, quote always attached. L4 is typed only on an owner utterance that names the item and the rung — that utterance is the signature — and the owner may always edit the file himself instead.
  9. Statuses are audited, in both directions. A periodic sweep asks of every status: what act does this claim, and where is that act recorded? Stale-low (work done, rung not raised) is the same defect as inflation.
  10. When the owner’s words are ambiguous, the rung is lower — and the ambiguity is resolved by asking, not left standing.