Three months ago I started building queuepilot with an AI coding agent doing most of the typing. Here’s what actually changed.
Plans first, diffs second
The single biggest workflow change: I stopped asking for features and started asking for plans. A plan I can react to in thirty seconds; a five-hundred-line diff I have to reverse-engineer takes twenty minutes and I miss things.
Small diffs review better
Agent-written code reviews best when the diff is small. One file at a time, one concern at a time. Review effort dropped by roughly half once I started enforcing this on myself as much as the tool.
Where it still struggles
Anything that requires holding invariants across files the agent hasn’t read yet — a shared retry budget, a lock ordering convention — still needs a human sanity check. I’ve settled into a rhythm where I own architecture and review, and the agent owns typing.