Autonomous agent control plane
A control plane that runs AI coding agents for a full working day without supervision and keeps each run inside its budget. Production code keeps shipping overnight, and model spend per change stays in cents.
Headline result
23 PRs in one 8-hour run, for $1.42
The goal
Ship production code changes overnight for cents in model spend, with engineers reviewing finished pull requests in the morning instead of watching agents all day.
The problem
Long coding-agent runs fail quietly. Agents mark work complete with checklist items still open, tests pass while the bug survives, features ship switched off, and context drifts after every reset. One false "complete" cost $238 before anyone caught it, and supervising every run by hand erases the savings.
What we did
Every run gets a time budget and a dollar cap, and every sprint works in its own git worktree, so parallel agents never write to the same checkout. Work moves through planning, architecture, implementation, QA and pull-request stages, and each stage has its own verifier.
A completion gate refuses "done" while any checklist item is open. Merge gates check behavior and front-end output before auto-merging. Circuit breakers halt new sprints until a person clears them, and a baseline check stops any run that starts on failing tests. Each sprint writes an event log and a report, and a narrator posts a one-line status every five minutes.
The results
- 189 pull requests opened across 245 unattended sprints in 103 recorded runs
- Longest verified run: 8.1 hours, 27 sprints, 23 pull requests, $1.42 in model costs
- 89 of 103 runs ended with the work queue empty. The other 14 stopped on a deadline (10), a failing baseline (3) or the kill switch (1)
- 2,727 automated tests cover the control plane itself
What it does for revenue
- Margins
- Model spend on the longest run averaged about 6 cents per pull request. Engineering hours go to reviewing finished work, which is where the judgment is.
- Speed to market
- The backlog keeps moving while the team is offline. An 8-hour overnight run produced 23 pull requests ready for review the next morning.
- Cost
- Hard dollar caps and circuit breakers stop a runaway run before it burns budget, the failure that cost $238 before the gates existed.