What problem does it solve? Agent loops can hit inputs their stop conditions never anticipated, spinning for thousands of iterations and generating massive bills before anyone notices. This Skill provides the layered enforcement — iteration caps, token and spend budgets, wall-clock limits, and kill switches — that bounds the damage when a loop's own stopping strategy fails. ## Core Features & Use Cases - Four-axis capping: Set independent limits on iterations, tokens/spend, wall-clock time, and expensive action classes (external API calls, message sends), with values derived from measured p95 distributions per task class. - Layered enforcement: Combine in-loop budget awareness, harness-level hard stops, platform-level API key spend caps, and billing alerts so no single failure point leaves the loop unbounded. - Graceful budget exits: Persist state, emit exit-labeled traces, and produce handoff reports when caps fire, so partial work survives and cap events don't become total losses. - Use Case: After an overnight agent run burns $3,400 on an unresolvable git edge case, use this Skill to rebuild with per-run caps, per-key daily spend limits, fleet spend-rate alarms, and a staging drill that verifies each layer fires in order. ## Quick Start Set hard iteration, spend, and time caps on my agent loop based on its measured p95 run distribution, with a harness-level kill switch and a budget-exit handoff report.
