Reviewing AI agent proposals without losing time
Review is the price of working safely with agents. This way of working keeps that price low.
A proposal from an agent comes in. Four hundred lines changed across eighteen files. The developer sighs, makes coffee and loses the rest of the morning.
It doesn't have to be that way. Review takes time, but how much time depends mainly on how you organise it.
Keep proposals small
You can review a fifty-line proposal in five minutes. A five-hundred-line proposal takes an hour, and you miss more. Ask the agent to split large tasks. One subject per proposal.
Start with the tests
First check whether the tests pass and whether new tests have been added. That already tells you a lot about the quality before you have read a single line of code. If the tests fail, send the proposal back without reading further.
Read the explanation, then the code
A good agent explains what it did and why. Read that first. If the reasoning doesn't hold up, you don't need to read the code.
Know what to look for
Check in particular:
- whether the change does what was asked, and nothing more,
- whether nothing changes in parts that were not in the task,
- whether no sensitive data or keys end up in the code,
- whether the naming and style match the rest of the code.
Measure it
For a few weeks, track how long review takes for each type of task. You will soon see which tasks really save time and which don't. Tasks that cost more review time than they save come off the list.
With us, review time is a standard part of every report. Time saved without review time is not honest time saved. Also read how we measure time saved.