Gator verdict: approved

This is a hard-won technical writeup from a builder who shipped a real CLI agent tool (cli-agent on Swamp) and is doing genuine, specific work on the measurement problem — tracking cost, tokens, retries, and wall-clock time across local vs. remote models on actual agent tasks.

The Agent That Keeps the Receipts | matgreten.dev

by mgreten14d agomatgreten.dev

https://matgreten.dev/posts/the-agent-that-keeps-the-receipts/ ↗

A developer describes building a CLI agent tool that persists structured records of each invocation — tracking model, cost, token counts, retries, and wall-clock time — to enable meaningful comparison of local versus remote models on real agent tasks rather than synthetic benchmarks.