A workflow for reviewing AI-generated code across many agents

Once you run more than a couple of AI coding agents, the constraint stops being how fast they write and becomes how fast you can review what they wrote. A single afternoon can produce a dozen diffs across several repos. Miss the review discipline and you'll ship confident-looking mistakes.

Here's a workflow that scales past two or three agents.

Read the diff before you read the explanation

Agents are persuasive. If you read the summary first, you'll anchor on the happy path the agent describes and skim the code that contradicts it. Flip the order: read the diff cold, form your own opinion, then read the agent's reasoning. Disagreements are where the bugs hide.

Triage in three buckets

For each diff, spend ten seconds sorting it:

1. Ship — small, obviously correct, tests present.
2. Question — plausible but touches something risky (auth, migrations, money, concurrency).
3. Redo — wrong shape; send it back with a sharper instruction rather than hand-fixing.

Most teams waste time hand-fixing category 3. It's usually faster to re-prompt with the specific constraint the agent missed.

Make the agents do the boring verification

Before you review, have each agent run its own tests, typecheck, and a self-review pass ("what could be wrong with this change?"). You're not outsourcing judgment — you're removing the mechanical failures so your attention goes to the substantive ones.

Keep sessions visually separated

Reviewing across ten identical terminal windows is how you approve repo A's diff thinking it was repo B's. Give each agent its own tiled panel with a name and color so the diff you're reading is unambiguously tied to a project. This is one of the reasons we tile sessions in Agent Tile: review accuracy drops fast when you can't tell windows apart.

Batch the broadcastable steps

"Run the tests," "pull main," "rebase onto the new API" — these are identical across agents. Broadcasting them to a group beats typing the same thing ten times, and it keeps every session at the same checkpoint so your review is comparing like with like.

The mindset

Reviewing AI-generated code is a different muscle than writing it. Half the skill is reading quickly and trusting your instinct when something smells off. The other half is refusing to rubber-stamp: an agent that's right 95% of the time still hands you a broken change every twentieth diff, and at fleet scale that's several per day.