What changes when coding agents open most of a project's pull requests?
A former engineering manager now runs a project where agents open the large majority of pull requests. Tickets became the spec and the test, nobody approves their own work, and cost per finished story replaced budget per head. For teams in the same position: what did you stop doing, and what did you have to start?
The work moves rather than shrinks. Writing tickets becomes the main job, because a ticket with clear acceptance criteria is what the agent builds against and what a checker compares the diff to. Vague tickets that used to be resolved in a conversation now turn into wrong code.
Keep the rule that nobody marks their own work done, and apply it to agents. The agent that wrote the change should not be the one declaring it finished: a separate step should read the CI output, compare the diff with the criteria and attach evidence such as test results or a screenshot.
Limit work in progress by how much the humans can follow, not by how many agents you can start, and make anything that needs a person visible in a blocked column instead of letting it wait quietly.
Track cost per finished task and watch the expensive tail rather than the average. The few tasks that go far over budget usually point to unclear tickets or an agent stuck in a loop.
Listings mentioned
- Verification before completion · skill by obraMakes the agent run checks and show the output before it claims a task is done.
- Requesting code review · skill by obraHands finished work to a separate reviewer, so the author never approves its own change.
Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.