The Habr author reports that AIWorks ran on its own repository for 24.5 hours and produced 63 completed pull requests, each independently checked. At peak, four workstreams ran in parallel; 18 pull requests were merged overnight. These figures describe one author's project, not a controlled productivity comparison.
The mechanism is separation of duties and constrained authority. The author says a fresh, read-only session checks the same commit and reruns tests; the supervisor pushes that verified commit and opens a pull request, but does not merge or deploy. That structure makes a checkable artifact—not an agent's assertion of completion—the handoff. The author also reports exhausting GitHub GraphQL quota and discovering a test that passed on Windows but failed on Linux because of filesystem case sensitivity. These failures show why repeatability and explicit publication semantics are part of the loop, not incidental cleanup.
Why now: coding agents increasingly invite longer unattended runs, making the cost of weak verification and ambiguous completion more visible. The evidence boundary is one self-reported case, and the article does not establish that the workflow generalizes or that every reported change had equal value. The first implication is that teams extending agent autonomy should define what artifact counts as complete and keep irreversible actions behind a separate approval boundary. The first watch condition is whether subsequent projects report independently reproducible outcomes, including review quality and rework, rather than pull-request volume alone.