DHH lets agents run for a day, then does the verify-and-merge pass himself
Oct 8 DHH on the Campfire rewrite: agents run for a day across competing implementations, he goes through every PR they opened, verifies the work, merges what's ready, then sends the agents back with architectural improvements for each one.
David Heinemeier Hansson posted an update on his experimental Campfire rewrite, which is running as several competing implementations. His loop: "I let the clankers run for a day, go through all the PRs that were submitted, verify the work, merge what was ready, and then ask it to bring all architectural improvements to each implementation." The C version is now in the lead.
It's a small post, but the workflow is the important part. The agents keep working without supervision for a whole day. The human doesn't steer during the run. He shows up once to verify and merge in a batch, then sets the next round of direction. Replies describe the same pattern ("letting agents run overnight and then doing a review and merge pass in the morning is the workflow that actually seems to stick").
It also shows where the always-on argument stops. The throughput comes from the agents, but deciding what counts as ready, and which architectural lessons carry across implementations, is still a daily human batch job. Steady agent work means the verification pass becomes the scheduled bottleneck, not that it goes away.