What they did
Jeremy Longshore runs a scheduled pipeline that writes and publishes a post a day to his startaitools.com blog, which is open source on GitHub. It ran for close to a year. Since the summer it grew from one writer agent to four passes: a writer, a classifier, a fact-checker and consistency checkers, with about 5,000 words of instructions plus some 38,000 words of reference docs. From September the default writer was MiniMax with Claude as the fallback. Each run has a 45-minute wall-clock ceiling. When runs began timing out, he worked with Claude Opus 5.5 to read the pipeline's own request logs and outage history, trimmed the writer's context, and ran a blind comparison of old and new posts.
How it went
One week of seven posts used about 205M tokens, nearly all cache re-reads, for roughly $105, about $15 a post. Three runs in a day timed out. A trimmed canary cut tokens from 104M to 30M and still missed the ceiling, while Opus runs on the fuller setup landed. The cause was four sequential passes on a slower default model, not request size. The outage history showed the pipeline had not run clean since June, including an 11-day silent abort and two lost posts. The blind comparison found the older one-pass setup wrote weaker posts, so he kept the checkers. The trimmed setup was merged, with its quality not yet measured.
Worth knowing
Under a fixed deadline, measure per-pass wall time and pass count, not just tokens. Before cutting agents to save time, run a blind read of the output: the cheaper pipeline can be the worse writer. The $15 figure includes failed runs and some of his own chat use.