One person. Twenty-one agents. Eighty-eight merged PRs in a week.
Rig Boss is the coordinator and dashboard DancingBits built around OpenRig, an open-source runtime for teams of AI agents. It makes OpenRig work with Jira and other spec systems, keeps the agents honest and busy for weeks at a time, and leaves you one job: making decisions.
- Agent “seats” that run in terminal sessions
- Teams of seats (“rigs”) defined in a YAML file
- A daemon and a work queue
- A terminal UI and a Slack connector
- Connector. Makes OpenRig work with Jira and other spec systems.
- Coordinator. Merge gate, claim checker, guard hooks, watchers and night mode.
- Dashboard. A board app for one-click decisions, a stage track per ticket, and a cost scorecard.
- Proof. Build and test queues on CI Macs, and scripted device screenshots per PR.
The problem a runtime does not solve
Starting twenty agents is easy. Keeping them useful for a month is not. Agents claim work they did not do. They wait hours for an event that never comes. They fight over the same files and the same machine. And a human who has to watch all of it stops being the scarce decision maker and becomes a babysitter.
A runtime gives you seats and a queue. It does not tell you where the work comes from, who checks it, when a human must decide, or whether the whole system is getting better or worse. Rig Boss is that layer.
How Rig Boss wraps OpenRig
Three rules Rig Boss enforces
Agents check agents
No agent merges its own work. A separate check seat gates each PR with an adversarial review. A claim checker tests each statement in a PR body against the diff, the tests and CI. Hooks refuse a malformed verdict or an unsafe command before it runs.
The human only decides
Every question for the human arrives in one inbox as a one-click card: approve, deny, or pick an option. You answer from the board, your phone’s home screen, or a thumbs-up in Slack. For routine choices, agents decide and tell you.
The system measures itself
A weekly scorecard counts merged PRs, round-one HOLD share, cost per PR, human wait time and idle seat hours. “Promises” watch the services the factory depends on, and go red when one breaks silently.
What is inside Rig Boss
Intake
- Spec-system connector
- You move a Jira ticket to one board column. Within two minutes a job queues it for the right seat. Moving it back withdraws the work.
- Mission cycle
- A product team reads the mission and goals and writes the tickets itself. The goals measure themselves.
- Night mode
- A night feeder starts only tickets marked safe for the night. A night watch keeps work moving every 15 minutes.
Coordinator
- Merge gate
- Adversarial review in rounds, sized to risk: two rounds for high-risk PRs, one for low-risk. Verdicts use a fixed format.
- Claim checker
- False claims in PR bodies were the top reason for a HOLD. This check now runs before the gate.
- Guard hooks and lane rules
- Hooks refuse unsafe merges, commands and device use. Lane rules keep two seats out of the same files.
- Headless merger
- Merges gated changes to Rig Boss’s own code every 15 minutes, with no human session.
Keep-alive
- Watchers
- Watch for decisions, stuck prompts, idle seats and a red main branch, and wake a seat only when its blocker may be gone.
- Seat refresh
- Replaces a seat near its context limit with a fresh one, and carries the work across.
- Factory overseer and promises
- An agent reads stall patterns, proposes fixes and runs a weekly retro. Health checks watch CI Macs, Xcode releases, backups and the daemon.
Dashboard and proof
- The board
- A web app, also on the iPhone home screen, with live activity, a stage track per ticket (build, PR, CI, gate, merge) and decision cards.
- Cost scorecard
- Reads every agent transcript and computes cost per ticket, per role and per PR.
- CI Macs and device proof
- Builds and tests go to two CI Macs through a queue. Scripted screenshots of the app in each device pose go on the PR as proof.
What broke, and what it taught
The failures shaped Rig Boss more than the plan did.
| What went wrong | What Rig Boss does now |
|---|---|
| Machine load hit 900 on 10 cores when three agents ran simulator tests at once. | One simulator run at a time, with tests moved to CI Macs. |
| Agents waited hours for an event that never came. | Watchers that wake a seat only when its blocker may be gone. |
| One sign-in expiry parked every seat at once. | Its own state on the board, and its own alarm. |
| False claims in PR bodies were the top reason for a HOLD. | The claim checker, run before the gate. |
Two design rules came out of it. No agent merges its own work. And green CI is not a merge signal, because green CI is exactly what a broken PR with a confident description looks like.
The numbers
From Rig Boss’s own weekly scorecard.
| Week of | PRs merged | Round-one HOLD share | Cost per merged PR |
|---|---|---|---|
| Oct 18 | 88 | 56% | $39.40 (partial) |
| Oct 25 | 72 | 52% | $36.30 |
| Nov 1 | 32 | 39% | $34.00 |
- The share of PRs held in their first gate round fell from 56% to 39% in two weeks.
- Cost per story point fell from $10.13 to $8.51 between the last two weeks. Story points were first recorded in the week of Oct 25.
- The median wait for a human decision was 5 to 18 minutes.
[Draft: confirm why the PR count fell in November. The brief assumes less queued work, not lower capacity.]
Lessons
- A step that fails silently looks the same as one that worked. Make failure loud, or make it impossible.
- Measure every rule. Each new rule gets a metric and a check date. Remove the ones that do not help.
- Spend human attention only on decisions. Everything a machine can check, a machine should check.
What this means for your business
The same three questions decide whether agents help any team: what checks the agent’s work, where a human must decide, and how you know the system is getting better. DancingBits helps you answer them as strategy, from experience of answering them in running code.
Talk about agent teamsRig Boss is built on OpenRig, an open-source project that DancingBits does not own.