I’m trying to find a better way to manage all the PRs my coding agents keep opening.
I’m thinking of splitting them into tiers:
- P0 / P1: I review and approve before merging.
- P2: The agent can merge on its own once checks pass.
Seems simple enough, but I still run into merge conflicts, and it’s hard to keep track of everything. I merge one PR, another needs updating, and a few others are still waiting for me to take a look.
I feel like I’m spending more time chasing PRs than I expected 😅
How are you guys handling this? Are you using merge queues, letting agents resolve conflicts and retry, or just limiting how many PRs they open at once?
Curious what’s actually working for you.
Top comments (2)
Capping concurrent agent PRs per area of the codebase helps more than a merge queue, since most of the conflicts come from two agents touching the same files.
Building on the comment about capping PRs per area: I'd treat the conflicts and the review pile as two separate problems.
Conflicts mostly come from how the work is handed out. Give each task an explicit list of paths it may touch, and don't start a second task whose paths overlap with an open one. Agents then rebase right before they mark a PR ready, so you never review something that's already stale. If you do that, the merge queue only has to catch real semantic conflicts and not every file collision.
For the tiers, don't let the agent pick its own risk level. Decide it from the diff: a small path-based config (docs, tests, generated files, dependency bumps = P2; auth, billing, migrations, infra = P0) applied by a labeling step or CODEOWNERS, so the label is there before you ever look at the PR. Keep P2 auto-merge limited to those categories and only on required checks, and have a way to revert quickly. Auto-merge is only as safe as the tests behind it, so anything the tests don't really cover stays a P1.
The "more time chasing than I expected" part is a flow problem. Put a ceiling on how many PRs can be waiting for you at once, and don't dispatch new agent tasks while you're at that ceiling. That way review capacity sets the pace, not the agents' output. A single list of PRs waiting on you, sorted by tier and age (for example
gh pr list --search "review-requested:@me"plus a label filter) beats checking each branch by hand.One last detail: ask the agent to put in the PR description what it changed, what it didn't touch, and how to roll it back. That turns a P1 review from reading the whole diff into checking a short list.