Building an AI Team for the Shop Floor

Part 1 of a series on multi-agent operations intelligence

It’s 2:47 PM on a Tuesday. Somewhere in a machining workshop, an operator finishes a part, picks up a clipboard, and writes a number into a small box on a paper card. Next to it, a one-line note in the local language about a tool change. By the end of the shift, that card joins dozens like it — a stack of paper that captures everything that happened in front of every machine in the building.

This data is precious. It tells you whether the shift met its plan, why a machine went down, what quality flagged for review. But it sits in a stack. It doesn’t talk to anybody. It doesn’t roll up. By the time the workshop owner sees it, three days have passed, and the shift is two cycles deep.

We’ve spent the last several weeks designing something to change that.

It is, in essence, a small team of AI agents that mirrors how a real factory leadership group works. There’s a Chief of Staff that produces morning briefings and end-of-day summaries. There’s a Production Manager, a Quality Manager, and a Maintenance Manager — each watching their domain, each surfacing what matters, each getting measurably better at their job over time. And there’s an Ingest Agent — the one that reads the hand-written cards from the floor and turns them into structured data the rest of the system can reason about.

It runs on off-the-shelf vision-capable language models, a Telegram-based interface (no special apps for floor staff to learn), and a small set of disciplines that have turned out to matter more than the technology choices: who can talk to whom, what gets validated where, how the system handles confusion gracefully, and how it improves over months rather than degrading or hallucinating with use.

The rollout is deliberately patient. We ship one agent at a time, each calibrated against real shop-floor conditions before the next comes online. The first phase is the one that decides whether the rest of the system has any data to reason about at all: card ingestion. If we can’t read the cards reliably, nothing else matters.

Over the coming weeks, we’ll publish more in this series — covering what each agent does, why it’s structured the way it is, the trade-offs we made, and what we’d warn anyone else trying to build something similar. Some posts will be short. Some will be opinionated. None will hand over the keys to the kingdom.

If you’re operating a manufacturing site and the gap between your data and your decisions feels too wide, we’d be happy to talk. The systems we build aren’t moonshots. They’re patient, instrumented, and reachable.