Verification, not coding, becomes the bottleneck for agent-heavy teams

On November 6, 2026, a technical note published in the AI section of DEV Community described what one team experienced once coding stopped being the slowest part of its delivery pipeline. The team ran about 25 engineering agents and used Claude Code with three Claude Ultra accounts. The agents, working alongside people, wrote most of the pull requests in the team's main repository.

As agents began to open pull requests in volume, verification became the real constraint. According to the note, a single full run of the test suite covered 30,000 tests and required 60 to 75 runner minutes. Because activity on an agent's draft and each subsequent push could each trigger a full run, with the merge queue kicking off another one, the pipeline had little room to shorten. In the short term, the team decided to reduce or pause agent activity, deferred merging of performance-sensitive code paths, and explored alternatives such as compiling and validating in self-hosted runners.

In terms of cost, the team reported GitHub Actions expenses of roughly $1600 per month against the $600 it paid for three Claude Ultra accounts. Queueing contributed as well: median wait time in the merge queue reached 12 to 14 minutes, so throughput hinged less on how fast agents wrote code than on how quickly CI approved it. With batched pull requests, individual ships, and a longer tail of authored code, the problem the note describes is the point where many teams currently stand: how to keep spending on testing managed while moving plenty of AI-generated code through CI.