Why your AI agent's retry loop is a silent tax

Your agent failed a task. It retried. It failed again. By the fourth attempt you have paid four times for work that produced nothing.

Retries feel free because nobody puts them on an invoice. They show up in your monthly token bill as background noise, mixed in with the runs that worked. When we audited real agent workloads, the pattern that stood out was not the cost per successful run. It was the cost of the attempts that never shipped anything.

Three numbers to track before you scale a workflow:

1. Retries per task, not per session. A session hides the loop.
2. Token spend on failed runs as its own line. It is rarely small.
3. The failure reason distribution. If half your retries are schema validation, that is a prompt problem wearing a retry costume.

The math that stings: at a 30% first-attempt failure rate, a workflow that "costs" $2 per task actually costs closer to $2.60 once you count the failed attempts feeding it. Teams that instrument this watch the rate climb as they add tools, because every new tool is a new way to fail.

Fix the failure, not the retry budget.

Full breakdown with audit numbers: https://pastagi.com/engineering/kv-cache-decode-cost/