2,700 agent-written PRs in seven months, and the slowest station on the line is the human approver
Erik Yu of Parall counted fifteen thousand of his own messages to the repo. Once a merge queue and self-review landed, "rebase" and "I merged it" fell to zero. What's left for people is a two-layer gate that skips reading code, and it is now the bottleneck.
Source: The Crofter Is the Bottleneck
Parall made its first commit in March. Seven months later, nearly 2,700 PRs have landed on main, every one written by agents, and Erik Yu's write-up opens with one of those agents messaging him: "All 8 user-feedback fix PRs are stuck in your approve queue." It was the third reminder. Feedback intake, diagnosis, the fix, tests and review had all been done by agents, and all eight stopped in the same place: waiting for him to click approve.
The most useful part is a count. Yu had an agent tally the 15,000-plus messages he sent the repo and chart a few phrases by month:
- "rebase" hit 69 in April. Writing "keep up with main" into the rules backfired as more agents ran: one merge forced every other branch to update, re-verify and wait for review again, and by then main had moved. A merge queue went live on August 23. In September he said "rebase" zero times.
- "/pr-review" went from 222 in April to about one a month once agents carried their own PRs through review.
- "I merged it" peaked at 116 in April and hasn't appeared since September. "Enqueue" replaced it.
- "continue" never went away: 94 in July, 46 in September.
Why it matters for code review
About a dozen changes land each day, and Parall doesn't gate them by reading code. In the first layer, machines review machines: CI, a code review tool for implementation detail, and an independent AI reviewer that sees only the product definition and design docs. The coding agent fixes what's right and pushes back with evidence on what isn't. That layer got smoother once the AI reviewer was told to judge only direction, because detail review kept spawning new mechanisms each round.
The second layer is a person looking at three things: the product-state doc diff (a 58-file, 2,600-line change came down to a few sentences), the running preview environment, and the agent's own account of the architecture. The architecture check is the one that won't automate. Yu asked some version of "is it clean?" more than thirty times and still has to say it out loud every time.
The complication for anyone betting on continuous agent loops is that the faster the machine layer got, the more the human layer stood out. Writing every correction back into the repo also has a cost. The root rules file grew from 66 lines to 483, started contradicting itself, and had to be cut three times, down to 171 lines.