By his own measurements, more than half of what Claude did in a session was drift away from what he asked for. His point for plants: refineries and labs will not hand their data to a model in someone else's cloud, so AI they trust has to run on a laptop, air-gapped, and give the same answer every time.
The agent has to work out which pump readings and maintenance revision to trust, find a qualified technician, reserve the right part, ask for supervisor review and draft a work order. A deterministic evaluator grades its tool calls and the final database state, with injected timeouts, rate limits, duplicate deliveries and a lost work-order response.
For reliability teams and vendors: prove a maintenance agent recovers from bad data and dropped responses, not just the happy path.
An agent can browse the node tree, search for tags like "Flow" and read live values, which is the slow part of a brownfield assessment. It ships with a water treatment plant simulator to practice on.
For SCADA and integration engineers: map what a server exposes in minutes before you build anything on top of it.
Open-source Python maintained at TU Berlin that runs dispatch and capacity planning, with more than 2,200 GitHub stars. Because it is plain Python, an agent can drive it to test how much battery a site really needs.
For power and data center developers: size onsite generation and storage with the same tool grid planners use.
