Crabhaus

What it is

Hard limits a harness enforces on what an autonomous run may consume - steps, tokens, time, money - plus explicit checks that define when the work is done. The run operates inside a contract: spend no more than this, stop when a verifiable finish line is reached, instead of looping until something breaks.

How it works

The limits live in orchestration code, outside the model: an iteration cap, token and cost meters with a kill threshold, step and run timeouts, and loop detection that trips on repeated identical tool calls. 'Done' is a check against the environment - tests pass, the file exists - never the model's own claim. On exhaustion, the harness checkpoints partial state and surfaces it to a human.

A concrete example

A nightly CI-repair agent runs with a 40-tool-call cap, a $5 ceiling, a 30-minute timeout, and 'done' defined as the test suite going green. One night it proposes the same failing patch twice; the loop detector halts the retry cycle and the agent posts its partial diagnosis instead of grinding the full budget on a fix that would never land.

Why it matters

Agents now run unattended for hours, and the default failure mode is not a crash but an open loop: unbounded spend, stale context, no signal that progress stopped. Budgets bound a bad run's blast radius in time and money; checkable stop conditions are what make long-horizon autonomy safe enough to ship.