How to review code written by multiple AI agents

Short answer: review in the reverse order you use with people. Start with what the change should not have touched, then check whether a test exists that would fail if this were wrong, and only then read the code. AI agents write code that looks right far more easily than they write code that is right, and line-by-line reading is exactly where that appearance fools you.

Why the order changes

With a colleague you trust the intent and review the execution. With an AI agent, intent is the uncertain part: it may have understood something else, solved an adjacent problem, or fixed the symptom. The code comes out clean either way, because writing clean code is what it does best.

Where do you start when reviewing an AI agent delivery?

The cheapest and most revealing question is "why did this file change?". A file you did not expect in the list is the most reliable sign the task was understood differently. It answers in ten seconds and saves you reading three hundred lines.

Then the test that proves it

Before reading the implementation, find the test. Two questions: does it exist, and would it fail if the implementation were wrong? The second is what matters. A test that passes with an empty implementation is common in AI agent output, because it was written to pass, not to prove.

Only then the code

And look for one specific thing: what was assumed without being stated. That the network answers, that the list is not empty, that the user has permission, that the date is in the right timezone. That is where agent output fails most, and it is invisible in a fluent read.

How do you review when three AI agents work in parallel?

When three AI agents work in parallel, history becomes a braid and reading commit by commit tells nobody's story. Review one whole front at a time, asking "is this task done?", and only then look at how the fronts interact.

That interaction is where the parallelism-specific defect lives: each front is right in isolation and together they disagree. No per-front review catches that, which is why running the full test suite on the combined result before merging is not optional.

Is it worth having an AI agent review another?

Yes, on one condition: it must be from another company. Two models from the same house give two versions of one opinion. From different companies they genuinely disagree, and the disagreement points at what to reread. It does not replace your review, it shortens it.

How not to become the bottleneck

Ask for smaller deliveries. A large task produces a large delivery, slow to check and hard to reject when wrong: you end up choosing between accepting a huge block and redoing everything. Three small fronts beat one big one because you approve incrementally.