What this is, and what it is not
The boss loop is a way of writing one prompt so that Claude does the checking instead of you. You hand it a task, you tell it to spawn a team of worker sub-agents to do the building, and then you add a third instruction that most people leave out: you appoint one agent as a strict boss whose only job is to tear the work apart. The workers build, the boss rejects, they rebuild, and the loop only ends when the boss scores the work as good.
"The boss loop" is our name for it. Nobody in the industry calls it that, and you will not find it in Anthropic's documentation. The published names for this family of technique are loop engineering and agent loops, and the specific two agent version of it is written up on this site as the Gauntlet Loop. If you have read that guide, this one is the fan out and the cost control: what changes when the builder is a TEAM rather than one agent, and what that does to your bill.
- The mechanism is real and it is what the Claude Code team themselves describe doing
- The name is ours
- The part almost nobody warns you about is the token cost, and it is the reason this guide exists
Why a builder cannot grade itself
Every agent setup that disappoints has the same shape. One model does the work, the same model is asked whether the work is good, and it says yes. That is not laziness and it is not something you can fix by asking more firmly.
- An agent that just produced something has the entire production in its context. It knows why every decision was made, so every decision looks justified. That is the definition of a bad reviewer.
- Without a reference, "is this good" has no answer. The model returns a plausible score because you asked for one, and that score moves between runs because nothing is anchoring it.
- A single agent optimises for finishing. It has a step list and it wants the list done, so "good enough" arrives when the steps run out rather than when the quality does.
The fix is structural. You need a reviewer that did not write the thing, holding it against something concrete, and the loop has to end on the reviewer's verdict rather than on the builder's step count.
Boris Cherny, who built Claude Code, on how he works now: "I don't prompt Claude anymore. I have loops running that prompt Claude and figure out what to do."
The three parts of the prompt
The prompt itself is short. Only the third part is unusual, and it is the one that gets dropped.
- The task, with a concrete example of what great looks like. Not an adjective. A real reference the model can open, read and measure itself against. "Make it clean" is not a bar. A file, a URL or a named product is.
- The fan out. Tell it to spawn a team of worker sub-agents and say how the work splits. Claude Code supports nested sub-agents, and Cherny has described running "armies of agents, with agents prompting agents in trees of thousands." In practice you want a handful, each owning a real slice.
- The boss. One agent, separate from the workers, that did not write the code, whose only job is to grade it against the bar in part one and send it back with specifics. The team's own term for this split is maker and checker.
The loop closes when the boss passes it. Not when the workers say they are done, and not when you get bored of reading.
Free, for your email
Get the full setup steps + the copy-paste setup prompt
The rest of this guide (4 more sections) opens right here. The setup prompt, one paste into Claude that sets it all up, comes by email only. Use your real address.
Unlocks: What the boss should actually be checking · The cost, and the order that fixes it · Running it without surprises · When not to use it
50,000+ follow where these guides come from. No spam, reply “unsubscribe” any time.