Agentic AI, industrial reality

The most useful industrial-agent piece I have read this month, and it names the real bottleneck: context, not model. "An agent is only ever as good as the context it can reach." Pieter’s move — an open, standards-based context layer and an external control harness for governance, both before the agent — is exactly the open-standards discipline I have argued for 25 years. On a plant, an agent has to know what sits upstream and downstream, and what its decision does to both. That is engineering, not prompting.

The Physical-AI version of "necessary, not sufficient," with a number that sticks: a picking robot at 95% lab success still fails around 50 times a day — a plant wants 99.9%. The gap is not one breakthrough away; it is distribution shift, latency, integration, safety certification and maintenance. This is the interface where software meets physical systems, and it is exactly there that the last one percent is the whole job. Demos scale; reliability is earned.

Older (2025) but foundational — and it travels straight into industry: another eval tool will not save the product; the scientific method and eval-driven development will. In regulated, high-consequence settings, "we cannot tell ahead of time when it is wrong" is the whole problem. The unglamorous process work — error analysis, held-out evals, monitoring — is the moat, not the model.

The boardroom confirmation of Pieter’s view from the plant floor: agents do not run on bolt-ons, they need re-plumbed infrastructure — data, context and governance as first-class citizens. Most enterprises layer agents on legacy processes and wonder why they stall. Same lesson, different altitude: the hard part is the substrate, not the demo.

The reality check in numbers: only about a quarter of organisations are scaling agents in even one function; the rest are experimenting or just layering AI on top. The gap between the narrative and the P&L is still wide — and it closes on exactly the unglamorous work this whole list is about: context, evals, infrastructure, trust.