How a Misplaced Retry Loop Drained 10 Million Tokens in One Night

A developer's prototype agent, built to summarize webhook payloads and post results to a channel, consumed its entire 10-million-token free-tier allowance overnight after processing just four jobs. The root cause was a retry mechanism placed at the wrong layer: when a tool call failed, the code retried the entire agent loop instead of just the failed tool, causing the full conversation history to be resent each time. With each retry, the previous failed attempt was appended to the context, compounding the token cost exponentially — a 2,000-token call ballooned to 200,000 tokens by the fourth retry. The issue was diagnosed using a custom trace logger that recorded token counts and conversation fingerprints per model call, ultimately leading to a four-line fix. The author notes that free-tier token ceilings, rather than being a drawback, can serve as an immediate diagnostic signal for runaway loops that paid APIs would only reveal on a monthly invoice.