Topics / Engineering practice
More agent-written code needs a better acceptance process.
If every generated change requires a senior engineer to reconstruct the system, faster code generation just creates a longer review queue.
Measure where the work goes.
Track time spent clarifying intent, finding dependencies, rerunning checks and resolving review comments. Separate that active work from elapsed queue time. Counting merged pull requests alone rewards small changes and can hide later rework.
Use that baseline to decide what to automate. Repeated evidence gathering is a different task from deciding whether a product promise should change. Keep the latter under explicit authority.
Give every change a common evidence structure.
Illustrative example
One agent promises retention. Another optimizes storage.
The first agent implements downloadable reports under an approved promise to retain each export for 30 days after completion. A later agent adds a seven-day cleanup job to reduce storage cost. Its tests confirm that old objects are deleted, while the report listing still offers their download links.
The second review needs the retained promise and evidence spanning the listing, object store and download path. Advance an export through the proposed cleanup boundary and check that the promised download remains available. Any shorter retention period needs an authorized product decision, not merely a successful cleanup test. The durations here are illustrative product choices.
- Retrieve approved retention and availability requirements before changing cleanup behavior.
- Expose the affected data flow, including metadata and stored objects.
- Run relevant checks across the changed lifecycle boundary on the proposed revision.
- Escalate missing evidence or a proposed amendment to the promise.
- Retain the decision and results so later storage changes inherit the same context.
Expand autonomy by evidence, not by enthusiasm.
Proof provides a shared requirement and evidence model for the agent and the reviewer. Start with one class of changes, then measure preparation effort, missed defects, unnecessary escalations and rework. Broader delegation is a policy decision after that trial, not a consequence of adding an agent.
A second model agreeing with the first can be useful review input. It cannot replace an observed result or authorize a change to the acceptance criteria.
Try it on your software
Find the repeated work in your review queue.
Bring a few recent agent-written PRs and one that required substantial human intervention.