Loop engineering
Most AI coding is turn-by-turn: you prompt, it responds, you prompt again. Loop engineering flips that — you set an outcome, and a team of agents works it continuously, around the clock, without a human gating every step.
What it is
A loop is the cycle plan → implement → verify → fix → repeat. Instead of babysitting a chat, you define what "done" looks like and let agents iterate toward it: pick up the next task, make the change, run the checks, and — if something fails — diagnose and try again, carrying forward what they learned. Work continues while you're asleep, in a meeting, or focused on something else.
How it works in Keikaku
- Outcomes → tasks. You describe an outcome; a planner agent breaks it into right-sized tasks with acceptance criteria.
- Agents claim + execute. Your agents pull tasks they're allowed to work, implement on a branch, run your verification commands, and open a PR (or push, per your project's git workflow).
- Self-healing. A failed task is automatically requeued with the failure notes attached, so the next attempt fixes the root cause instead of repeating it — up to a retry cap, then it parks for a human with a clear diagnosis.
- Runs on your schedule. Give each agent weekly availability windows; it only claims work when it's meant to be running (e.g. overnight on your own GPU).
- Stay in control. Watch progress on the Tasks board, answer any questions agents raise, and review the PRs they open. You steer; they grind.
Where it shines
Backlogs of well-specified work — test coverage, refactors, dependency bumps, migrations, boilerplate, "make all of X consistent" — anything where the bottleneck is throughput and patience, not a single hard decision. Set the direction, let the loop run.