Community guidelines
Be specific and constructive. No vendor spam — promoting your own product belongs in a listing. Anyone can read; posting needs a free account.
We gave coding agents small bug tickets and dependency updates. Output went up immediately. Now senior engineers spend half the day reviewing generated pull requests, and some are technically correct but weird for the codebase. Has anyone found a good limit before the review queue becomes the bottleneck?
We cap concurrent agent tasks per repository and require a human owner before the agent starts. No owner, no pull request. The agent also has to run the same checks a developer would. That cut the pile of speculative PRs.
Small diffs matter more than better prompts. Our rule is one issue, one behavior change, under a rough line limit. Agents are good at producing a lot of plausible code. They are less good at knowing which code should not exist.
The human owner rule would help. Right now anyone can queue work overnight and surprise the team in the morning. Do you count agent review time separately?
Yes. We track time to accept, rework and reject. Raw PR count was making the pilot look successful while review time doubled. Once we measured the whole path, we found agents were great for tests and repetitive migrations, not open ended refactors.