Humans and agents

A knowledge base built for a person typing a question into a search box will be rebuilt within a year, because its main consumer will become a process: an agent that retrieves, reasons, retrieves again, and acts.

That consumer needs things a search box does not. It needs the catalog exposed as a set of tools with descriptions good enough that the agent can choose the right knowledge product for a sub-question. It needs retrieval results with structure — scores, lineage, freshness signals, conflict flags — rather than text alone. It needs citations it can carry forward into its own output. It needs the option to escalate from summary to chunk to source, and a budget it must respect while doing so. And it needs the knowledge base to be honest about what it does not know, because an agent that receives a confident wrong answer will act on it.

The same knowledge base should answer a person's natural-language question through the same retrieval layer, with the same freshness and grounding guarantees. Building two knowledge bases — one for chat, one for agents — is the mistake to avoid. Build one, with a retrieval contract rich enough for the agent, and the human interface is a simplification of it.