The factory demon died overnight.
No warning, no error. The background process that watched the ticket queue simply stopped running while everyone was away. By morning, tickets sat idle, and the failure had no explanation yet.
So the developer opened a new GitHub issue: "factory demon dies overnight." He typed only a rough description of what he'd noticed, no diagnosis, no fix, just the problem as he understood it. Then he added a label: factory ready for spec.
An agent picked it up immediately. It spawned an isolated git worktree, read through the codebase, and began investigating, inspecting logs, tracing behavior, hunting for the root cause. Minutes later, it had reproduced the bug and rewritten the ticket entirely: clear findings, a real diagnosis, acceptance criteria a human or another agent could act on.
Satisfied, the developer added a second label: factory ready to implement. A different agent took over, spawning its own worktree. It wrote the fix. It ran the tests. It caught issues, adjusted, and ran them again, working uninterrupted while the developer did nothing but wait.
About thirty-five minutes later, the work was done. A pull request sat ready, complete with an automated thumbs-up from Codex confirming the change was sound. The developer reviewed it, approved it, and merged it in.
The factory demon came back to life, quietly resuming its watch over the queue, waiting for the next ticket, ready to run the same cycle again.
source video: https://www.youtube.com/watch?v=AbpyqAfxZ8c