Researchers wired large language models into a simulated factory and let them run it, faults included.
Engineers paired each module of a simulated six-module hexagonal factory with its own LLM-based agent, connecting them through standardized MCP tool servers that expose machine functions via OPC UA, with agents talking to each other over MQTT and receiving live updates on factory state. The team tested three ways of organizing the agents: a single monolithic agent running everything, a peer-to-peer setup where agents coordinate as equals, and an orchestrator that directs the others. Across nine production tasks of rising difficulty, the monolithic and peer-to-peer setups both solved 93% of problems on average. The orchestrator version stood out on one specific failure: it correctly diagnosed and rerouted around a silent conveyor-belt fault in all ten test runs, something the researchers never explicitly programmed it to handle.
Factories making smaller, more customized batches need machines that can be reprogrammed quickly, and today that mostly means manual engineering work. This research suggests LLM agents could handle both the offline job of writing production sequences and the online job of noticing when something has gone wrong without being told in advance what wrong looks like. That is a different pitch than typical industrial automation, which usually assumes failure modes are known ahead of time.
It is still a simulation of one hexagonal test factory, not a working production line, so treat the 93% solve rate as a lab result, not a guarantee it survives contact with a real shop floor.