Bound how many times an AI agent can retry or call a tool, and design it to break out once a result is good enough.
Hard to implement
8 High impact score
Agentic workflows can spiral into many small tool calls chasing a single task, each one adding its own compute and context cost without a cap, a single user request can silently balloon into dozens of round trips. Set sensible limits on iteration and retry counts, and build in a clear exit condition so an agent recognises and stops once it has a good-enough result, rather than continuing to refine indefinitely on diminishing returns. It’s the same principle as the Circuit Breaker pattern in 7.4: fail or break out fast rather than exhausting resources against a problem that isn’t resolving.