$0.0068 to Kill a $0.26 Retry Loop

When an AI coding agent runs a bad command, it doesn't just fail. It fails, reads the error, reasons about why it failed, generates a new command, and tries again. Every step in that chain costs inference. The cost compounds with each retry.

The anatomy of a retry loop

Here's what happens when an AI agent runs a command that doesn't work:

1. Generation — the model generates a command ($0.03-0.10 in inference)
2. Execution — the command runs and fails (3-10 seconds wasted)
3. Error ingestion — the error output gets added to the context window
4. Reasoning — the model processes the error and decides what to try next ($0.03-0.10)
5. Retry — back to step 1, but with a larger context window, so each subsequent call costs more

A single bad command typically triggers 3-5 cycles before the agent either finds a working alternative or gives up. The context window grows with each error, so each retry costs slightly more than the last.

The math

One Check call costs $0.0068 AUD. It validates the command before execution. If the command is invalid, it gets blocked. No error in the context. No retry. No compounding cost.

One retry chain — the failed command, the error ingestion, the reasoning, and 3-5 retry attempts — costs roughly $0.26 on average for API-priced models. That's a conservative estimate. Complex commands with large error outputs push this higher.

Every Check call that blocks a bad command saves that entire chain. $0.0068 in, $0.26 saved. That's a 38x return on a single call.

At scale

A typical AI agent session generates dozens of commands. In autonomous coding sessions, the failure rate on unvalidated commands can be significant — wrong flags, missing binaries, unreachable targets, bad syntax. Each one triggers a retry chain.

If Check validates every command before execution, the bad ones get blocked instantly. The agent never sees the error, so it never retries. The context window stays clean. The session runs faster and cheaper.

Install

Check runs as a hook in your IDE. Every command the AI generates passes through Check before it reaches the shell. If it's valid, it runs. If it's not, it's blocked. The agent moves on instead of retrying.