Run a client project where coding agents open most of the pull requests

The poster, a Redditor who helped grow an engineering org at a services company from 30 to 150 people, spent this year running a client project where coding agents write most of the code. Work is driven by tickets whose acceptance criteria double as the spec. No agent signs off its own work: one step reads CI, another compares the diff with the criteria, and a third clicks through staging and takes a screenshot for each criterion. Any of those steps can send a card back. Up to 15 stories run at once, and anything that needs a person lands in a Blocked column on the board. People own the requirements and the final sign-off with the client.

Ten weeks in, the project had 1,575 pull requests, 1,363 of them opened by agents. The poster tracks cost per finished story rather than per head: the median is $36, but one story in ten goes past $147, and that tail is what they watch. Code review is about 11% of agent spend, and people no longer read every diff. Most of the poster's own time now goes into writing good tickets. The work-in-progress limit is set by how much the humans can follow, not by how many agents are available.

The poster does not name the agent tool or the models, and the thread's comments were not captured. They say the part they have not solved is growing the humans on the team: an agent only improves when its instructions do, so feedback that used to turn seniors into team leads has no obvious replacement.