Budgets and Loop Control for AI Agents: Caps, Stop Conditions and Graceful Failure

Agents can loop, wander and overspend because the model decides when to stop. Budgets enforced by orchestration code (steps, tokens, cost, wall-clock time), repeated-action detection, failure-specific retry rules and explicit terminal states keep cost predictable and failures graceful. This guide covers each control, how to size it, and the causes of loops worth fixing first.