The Intent Audit

What did your agents decide last month?

Five questions against your agent-shipped PRs. One PR to learn the method. Five for a useful signal. Thirty for a real baseline — or do it live with the founder.

Your agents' hardest calls are rarely obvious mistakes. They are reasonable local choices that can become precedent for every developer and agent that follows. These questions show which ones did — and whether anyone owns them.

The five questions

Pick one agent-shipped PR from the last month. Ask these, in order. Score it honestly. Nothing to install, nothing to share — you can run this on one PR before your coffee cools.

  1. 1. What local call did this PR make?

    Not the code — the locally reasonable choice inside it. An interface shape, a latency tradeoff, an exception, an ownership boundary, an interpretation of the requirement.

  2. 2. What standing decision applied — and did the agent have it?

    Your team may already have decided something that governed this work. Was it in front of the agent, with its scope and rationale, or did the agent have to infer it from nearby code?

  3. 3. Did the work support that decision, challenge it, or need an exception?

    A mismatch is not automatically an error. The implementation may expose a stale assumption, a better path, or a legitimate local exception.

  4. 4. What became standing after this PR — and who owned that transition?

    Was the prior decision reaffirmed, the local choice adopted, or an exception made explicit? If there is no answer, the code set precedent by default.

  5. 5. What related work should now build from — or reconsider?

    Name the agents, workstreams, or code paths affected by the answer. If nobody can, the next run will learn the precedent from code and quietly decide again.

How to score

A question passes when the answer is explicit and evidenced, not inferred after the fact from the diff. The number to watch is PRs where question one has an answer and question four does not: a local call exists, but nobody explicitly decided whether it should hold beyond the change.

Whatever your number is, it was invisible this morning. It is also why review does not scale: every unresolved transition becomes precedent in code or gets re-decided by someone senior later.

The first time we ran this on our own repo, it wasn't flattering.

What the questions add up to

These five questions are the shape of the loop. Agents start from the decisions that apply. The work can support them or push back. Proposed change stays distinct from what currently holds. An owner resolves the transition, and affected work builds from the answer.

Local choices stop becoming shared assumptions by accident.

Do it live with the founder

Ben Werther, Igni's founder, runs every session personally — it's how the early Intent Audit baseline gets built. Bring one recent agent-shipped PR — or the teammate who knows it best. You drive on screen — no repo access, no file upload. You control what is shown.

It's a 30-minute working session. Teams that want the real number go on to score their last thirty with the same method — everyone starts with one. We're running a first cohort of ten, for teams whose agents already do production-bound work.

Before

Pick one recent agent-shipped PR you know well — or bring the teammate who does. Nothing to install, nothing to connect.

During

30 minutes. You drive, on your screen. We work the five questions through the PR together.

After

You leave with your number, the specific gaps, and what we'd do about each — in a short written follow-up. Your findings stay yours; if you opt in, anonymized aggregate scores join the early study.

We're building an early picture of agent-built precedent: when local calls become shared assumptions, whether the transition was explicit, and what related work inherits.

Request a live Intent Audit

Fair questions

Isn't this what our docs and rules files already do?

A spec can stay current and still leave the hard questions unanswered: who may change a standing decision, where it applies, and what other work must now be reconsidered. The five questions surface exactly that gap.

You want to read our PRs?

You drive, on your screen — no repo access, no file upload. Nothing leaves unless you opt in to the study — and then only anonymized aggregate scores you've reviewed first.

Is this a sales pitch?

I built Igni — the audit is how we're measuring the problem it exists for. The session runs without Igni: nothing to install, nothing to adopt. If the findings make you want the fix, that's a separate conversation, and it's yours to start.

Request a live Intent Audit

The form takes a minute. Tell us where to reach you and we'll arrange the session. If a note brought you here, reply to it instead.