Token-min'ing

I've had a Claude goal running locally for almost two days straight, basically doing lightweight supervision of a crawl script.

Crawls are unreliable in all kinds of boring ways. If I just kept rerunning a shell script and handed every error back to Claude, there'd be a ton of back and forth. Instead, the goal stays alive, watches the process, and only does something when it needs to.

The surprising part is how little usage that takes. I'm still only around 10% of my weekly limit on the $100 5x plan.

There's something here that feels like token-min'ing: let the cheap deterministic process do almost everything, then keep a small amount of intelligence sitting on top of it for days at a time.

The agent doesn't need to be constantly thinking to be constantly useful.