Rework and Concurrency

Rework and concurrency

Sections 5 to 6 of the method. Fixes go back to the agent that wrote the change. At most four changes in flight.

5. Send work back to the agent that did it

A review finding, a failed build or a conflict goes back to the agent that wrote the change, not to a new one. It still holds the context. A fresh agent re-reads the codebase to reach the same conclusion, and pays for the reading.

I measured this rather than assuming it. Resuming a drafter to make a two-line fix its reviewer had asked for cost about 4 thousand tokens. The same agent's first pass, which included reading the repository, had cost 154 thousand. The saving is not marginal.

There is a second reason, which matters more than the money. The agent that wrote the change knows what it already checked. A new one does not, so it either re-checks everything or, more often, assumes the first agent got the basics right and looks only at what you pointed it at.

6. Four changes in flight

The merge queue takes one change at a time and re-tests each against everything before it. Ten in flight means every later one gets brought up to date, re-tested, and sometimes repaired. That rework costs more than the parallelism saved.

---

Previous: Scripts, not instructions | Home | Next: Measure every batch