joseph.dattilo// lansing · mi

← All writing

AI & agent fleets

Running AI coding agents against production repositories, every working day. The operating discipline — work queues, scoped credentials, transport-level enforcement, human gates — and the longer history behind it: production natural-language processing since 2018, before large language models existed. This is the thinking that became the FleetHarbor suite.

If you only read one thing, read the fleet essay — it is the distilled version of the discipline I apply every working day: the queue is the system of record, agents never hold real credentials, enforcement lives at the git transport layer instead of in the prompt, and humans keep the gates that matter — what gets built, what ships, what the fleet may touch. The 2018 essay is the receipts behind the claim in its title: what operating machine learning in production actually taught me (the model is a component, evaluation beats intuition, failure is a distribution), and which of those lessons LLMs changed — fewer than you would think. Everything here comes from systems I actually run — the same fleet that builds the FleetHarbor suite ships my own products every working day. More lands here as that work keeps throwing off lessons worth writing down.

The posts