How AI-written code gets proven before it counts

An agent will tell you it’s finished as soon as the code compiles, and code that compiles can still put a blank white screen in front of a customer. If your reviews can’t keep up with what your agents write, the answer is evidence: every change arrives with proof that it works, so reviewers check the proof instead of guessing. This is the standard we run on our own builds, and a smaller version can become part of how your team reviews AI-written code, the Review handoff in the Four Handoffs.

Risky plans get challenged before they run

For changes to architecture, infrastructure or security, the plan is written and challenged before any code changes. Three reviewer agents read it with narrow briefs, covering the evidence behind every claim, security and access, and second-order effects or decisions that can’t be undone. Must-fix findings go into the plan first, and anything put off is logged with a reason.

Done means someone looked at the page

The build runs as a loop of plan, build, verify, test, review and fix. It starts by playing back what was asked for as a checklist, and that list becomes the contract for everything after it. Before anything counts as done, the loop requires opening every affected page in a real browser, signing in, and checking that the page loads, the console is clean, the new feature is actually there and the buttons work when clicked. “It looks correct” doesn’t count, so there has to be a screenshot, checked against a list of what should be on it. When a test fails, the code gets fixed and the test is left alone unless the test itself is wrong, and work that won’t come right goes back to a person with its history.

What counts as proof

Proof is concrete: a file and line with the code quoted, test output with timestamps, logs, database snapshots and diffs, and never an agent’s “I checked and…”. When a reviewer agent checked a change to the math in Debt Hacker, the fintech app I invented and we built at Signal Scout with AI agents and a data pipeline, it wrote verification scripts of its own, checked the new calculation across 504 cases in a separate session from the agent that wrote the change, and backed each of its conclusions by running code rather than by reading it.

Stopping a risky edit before it happens

Some mistakes are cheaper to block than to catch. On one codebase a short script runs before every edit an agent makes and can refuse it outright, for example a write to a file that holds credentials or a password typed straight into the code, and it points the agent to environment variables or a secret manager instead.

A smaller version for your team

On your team it starts with two pieces: a short note on every pull request that says what was run and what couldn’t be checked, so reviewers can see what was proven without running it again, and a look at the change in a real browser before it counts as done. Nothing counts on an agent’s word, and every rule the team adds needs your yes, or the yes of the owner you name. The guardrails around it are on the engineering side. Walk me through one recent pull request in a thirty-minute call, and we’ll look at what it would have taken to prove it.

Read next

Does your plan match your code?

How I build with AI