You can't run lights-out on tribal knowledge

Every vendor in AI is selling the same picture right now: the dark factory. Agents that build, check, and ship while nobody watches. Manufacturing got there first — FANUC’s robots build other robots, unattended, for up to 30 days at a stretch.

But here’s the part of the manufacturing story that gets skipped. Lights-out didn’t happen because the robots got smart. It happened because engineers spent decades making the inputs boring: every part arriving in a known form, every check machine-readable, every machine reporting its state. Autonomy was the result of that work, not a substitute for it.

Service teams are trying to skip the boring part. They point an agent at the knowledge it needs — and that knowledge is tribal. It lives in chat threads, personal notebooks, shared drives, and the heads of whoever’s been there longest. Worse, it’s duplicated and quietly contradictory: two articles, two different answers, both “official.”

OpenHands, writing about what software can learn from lights-out manufacturing, names three prerequisites. Translated for a service team:

Standardized inputs. One question, one answer, in one place. Tribal knowledge is the opposite of a standardized input — it’s variability the agent has to absorb with judgment.

Deterministic verification. A check that can tell you the answer was right without a human squinting at it. If “correct” requires someone senior to nod, you don’t have verification. You have supervision with extra steps.

Continuous instrumentation. If your agent answers 500 questions overnight, what did it say? Without an execution record, you’re not running lights-out. You’re running blind.

The line from that piece I’ve been chewing on all week: “The more important correctness depends on subjective judgment, the harder the workflow is to run unattended.” Every tribal answer forces a judgment call. Judgment calls are the ceiling on autonomy — and no model upgrade raises that ceiling. Only the wrapper does: deterministic checks around a probabilistic core.

The takeaway: you don’t earn lights-out by buying a better model. You earn it by making your knowledge boring.

One thing to try this week: pick five questions your team answers from tribal knowledge — the ones that live in chat threads and people’s heads. Ask your agent each one. Every blended, hedged, or contradictory answer marks a place where the lights are still on. That’s not a model problem. That’s a work item.

The Simon test: what permanently changed this week that improves your system’s capacity to adapt?