Our review agent reads the task before the diff
A person asks for a button that exports a report as a PDF. The agent builds a button. It exports a CSV file. The code is clean. The tests pass.
A reviewer that reads only the diff sees good code. It has no way to know the person asked for a PDF.
What a diff shows
A diff shows lines that changed. It does not show why. The reason lives in the task. It is in the description, the plan and the list of things that must be true at the end.
A reviewer without that context can check style and obvious bugs. It cannot check the one thing that matters most. Did the change do what the person asked?
What the review agent gets
Every stage of a task starts with a brief. The brief holds each task field that has a value. The order fits the stage. The description comes first. The acceptance checklist comes next, ahead of the other fields. The plan follows.
The checklist matters most. Each item names the stage that proves it. The brief tells the agent which items are due now. So the review agent knows which promises it must check. Done means proven explains these proof stages.
What the review agent must do
The review prompt gives the agent a short job:
- Read the diff of the merge request.
- Read the acceptance criteria.
- Pick one action: approve or reject.
The rules are strict. The agent cannot approve while an item due at review is unticked. A later item needs a reason if the agent defers it.
A reject goes back to the code stage. The agent must post a comment first. The comment says what must change. The next code session reads that comment.
One more rule guards the scope. The agent must reject a plan or request that shrinks the task’s goal to fit the repos it has. The fix is to attach the missing repo. The task keeps its full goal.
Why we built review into the task
We could plug a separate review product into this stage. We did not. Such a product sees the diff and little else. Our review stage sits inside the task. It reads the same words the person wrote.
What it costs and what it catches
Review costs money on every story, and we do not hide that. In our anatomy of an $8 feature, review cost $1.11. That is 14.7% of the spend for stories near the median total. The data comes from a week in July 2026.
Review does catch things. In the week of 21 to 27 September 2026, we shipped 355 stories. Of these, 126 (35.5%) went back to an earlier stage before the merge. Review or merge sent them back. The same week’s data shows that these stories cost more than clean ones.
We cannot say how many of the 126 were a wrong goal and how many were a code fault. We did not measure that split.
Where it stops
The review agent is an agent. It can misread a task, as the author can. An agent that misread the task often misreads it again.
If you need a person to be accountable, add an approval gate. Who approves the agent’s work shows how. The review guide covers what a person does at that stage.
Review also cannot fix a weak task. If the criteria say “make it better”, the agent has nothing to check.
If you do not write code
You can still control review. Write acceptance criteria that a stranger can test. Say what the user sees and what happens next. Name one outcome for each line.
Our guide on writing a brief shows how. The brief checker scores a draft before you submit it. Try it on your next idea. The review agent will hold the work to what you wrote.