The agent looped on a broken tool all weekend.

A bigger budget wouldn't have helped — a static cap just delays the same burn. Same broken agent, run three ways through the Guardrail gateway. Fake money, real logic.

No guardrail

$0.00
0 calls
idle
Broken agent loops on the same failing tool. Nothing stops it.

Guardrail on

$0.00
0 calls
idle
Same agent, same bug. The breaker watches per-run.

Legit long task · guardrail on

$0.00
0 calls
idle
40 distinct steps. Proves the breaker doesn't punish real work.

The breaker hashes each tool call and cuts the run when the same signature keeps failing — so it tells a stuck loop from a long, legitimate task. It runs on the request path, per run, not a day late on the invoice. Open source: github.com/amih/guardrail · part of the AI cost + safety playbook at smallestbusiness.com.