What they did
Vojta Biberle's team at Keboola still had part of Keboola Connection on Zend Framework 1, with the rest on Symfony. He pointed an existing agent setup at that one goal with a two-sentence brief: remove as much Zend as possible by Friday, work in worktrees, use the repo's skills, and do not open too many PRs at once. The setup held the rest. A migration skill came with a map of every place the modern code still depended on the old code, generated once with AI and kept current by the agent and a checking script. Other skills took a task from Linear to just before merge, ran end-to-end tests, watched PRs and review threads, and set the gates for which changes a reviewer bot could approve and which needed a person. A main Claude Code session acted as organizer, starting one subagent per git worktree and checking each met the merge conditions.
How it went
In about 22 hours, 13 PRs merged and 3 were open, with nothing stuck. Median time from open to merge was about two hours, the fastest 57 minutes; overnight PRs waited 10 to 13 hours for review. The changes were small, and 56 dependencies on the old code were removed. The slowdowns were the E2E queue (30 to 90 minutes per run) and waiting for human approvals. The agent twice started E2E before the PR image was built and once used the wrong filter, and fixed both itself. Once it picked work from a stale local main that was already merged, wasting only agent time. About 70 remaining actions still need a person.
Worth knowing
The short prompt only works because the harness already encodes the process: the dependency map, the merge gates and the rule that the agent never re-pings a reviewer. Have the agent pick work from the remote main branch, not a local copy. The author is describing his own company's setup on his own blog.