Dual-Agent PR Loop¶
The setup Ryan Lopopolo's team runs for 0% human-reviewed code pre-merge: one Codex agent authors and pushes a PR; on each push, a reviewer agent fires and comments; the author agent is explicitly instructed to acknowledge and respond. Then two tuning tricks that matter more than the loop itself:
- Initial failure mode: the author agent was too deferential — "willing to be bullied by the PR reviewer" — leading to non-convergence.
- Fix: both prompts were given more optionality. The reviewer was told to bias toward merging and not surface issues that didn't block shipping. The author got license to push back when appropriate.
Merge conflicts and worktree pain (cursor has written off worktrees as "too many merge conflicts, too unintuitive") are resolved by the model: "the models are really great at resolving merge conflicts. To get to a state where I'm not synchronously in the loop in my terminal, I almost don't care that there are merge [conflicts]."
This is the PR-as-hub-and-spoke-broadcast pattern from the original talk made concrete — throughput over blocking review. It sits downstream of tokens-need-critique-loop — critic + author in dialogue — and is a direct extension of harness-engineering.