Claude! Stop Burning Tokens on Your Agent's Tool Output!

A Two-Stage Curator That Pays for Itself I watched Claude Code feed 108,894 bytes of seq 1 20000 back into its own context window. That output contained 20,000 integers. No errors. No signal. No insight. Just counting. And yet the system still had to tokenize it, send it back to the model, and bill for it. This is not an edge case. It is the default failure mode of agent tooling. Tools produce output. The output goes back to the model. You pay for it. Logs, test runs, ps listings, git history, b