If we want them to do Knowledge Work, design them as Knowledge Agents — Benjamin Clavié, Mixedbread
Knowledge agents need intent decomposition, specialist researchers, and multiple retrieval primitives; copying a coding-agent loop fails when evidence is ambiguous and context-dependent.
Clavié argues that non-code knowledge lacks durable identifiers and explicit definitions: a phrase such as “30 days” can represent several unrelated obligations. Search therefore begins with intent and conditional context, not a grep-ready target.
Design the main agent like a professional lead: decompose the question, send bounded research to **specialist subagents**, and combine their memos. Give the harness multiple primitives—**BM25, semantic search, and multimodal PDF search**—and teach it when each applies.
Clavié argues that non-code knowledge lacks durable identifiers and explicit definitions: a phrase such as “30 days” can represent several unrelated obligations. Search therefore begins with intent and conditional context, not a grep-ready target. Design the main agent like a professional lead: decompose the question, send bounded research to **specialist subagents**, and combine their memos. Give the harness multiple primitives—**BM25, semantic search, and multimodal PDF search**—and teach it when each applies. The cited benchmark result says improved search orchestration reduced the human-agent mistake gap by **40%**, but the talk omits enough setup detail to generalize that number. Better orchestration also cannot recover evidence absent from the corpus or retrieval tools.
This sharpens knowledge-agent design around ambiguity: the lead agent must infer intent, decompose research, and choose among retrieval modes rather than merely distribute queries. It supports bounded specialist subagents while conflicting with fixed multi-agent chains that lose shared context. The reported mistake-gap reduction is evidence for orchestration, but remains too underspecified to generalize.