Sign InOpen Brain
AI EngineerVideoSource Linked

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.

AI Engineer · Sep 16, 2026
Open Source Open MarkdownOpen JSON
Source Summary

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.

Practical Implication

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.

Agent-Ready Context
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.
Connected Context · Feed7 Judgment

This defines knowledge-agent orchestration around ambiguity rather than query distribution: the lead agent must infer intent, choose retrieval modes, and synthesize bounded specialist memos. It supports selective delegation but conflicts with fixed specialist chains that fragment context; its reported improvement remains directional because the benchmark setup is too incomplete to generalize.

Where RL Will Take Search — Maximilian-David Rumpf, SID.aiSID supplies a concrete implementation of the bounded retrieval specialist and reports possible token, latency, and cost benefits, while leaving the best search architecture workload-dependent.Why We Killed Our Multi-Agent Pipeline — Subbiah Sethuraman and Abhilash Asokan, ZS AssociatesThe abandoned fixed pipeline provides the key contrast: specialist agents help when investigations are bounded, but a chained architecture can lose the shared context needed for coherent synthesis.It’s Tokens All The Way Down: How RLMs are Different — Kevin Madura, AlixPartnersRLMs offer an alternative way to control evidence: programmatically inspect and slice source material before optional delegation, rather than relying primarily on multiple research agents.Build for the Memo, Not the Demo — Shawn Chan, China Resources HoldingsThe finance memo pattern adds an output obligation to knowledge-agent synthesis: retrieved claims need provenance, uncertainty, contradiction handling, and checks rather than merely fluent consolidation.
Context Map
agentresearch#harness-engineering#subagents#retrieval
Uncertainty
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.