Notes from the team building Amybot.

How we ship a credit-based ledger that survives partial failures, refunds, and webhook re-deliveries — without losing a single cent.

Seat licenses force teams to over-buy. Per-call pricing punishes experimentation. Credits are the only model that fits how agents actually get used.

What we measured running Amy with aggressive prompt caching: where it pays off, where it backfires, and the rule of thumb we use now.

Running an agent's browser actions next to production data is a footgun. Here's the isolation model we landed on.

Traces work great until your span outlives your trace exporter. Notes from instrumenting eight-hour agent jobs.

Cron-style triggers, idempotent dispatch, and the surprising amount of work that lives between 'fire' and 'run'.

We replaced a Redis-based mutex with a single Postgres advisory lock and removed an entire class of races.

Token buckets in Postgres, no Redis, and a hot-key strategy that doesn't pretend hot keys won't exist.