If we want them to do Knowledge Work, design them as Knowledge Agents — Benjamin Clavié, Mixedbread
Source
youtube.com
Author
AI Engineer
Date
Context
A knowledge-work AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → cannot always find the right evidence by matching the right phrase. In a contract, “30 days” could describe a deadline, grace period or retention rule. This knowledge-agent design proposal argues that code offers unusually explicit search cues, such as identifiers and file paths, while other documents require the agent to recover meaning from context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition →.
The proposed response combines better retrieval tools with searcher agents that return short memos. Those agents take on the work of finding and interpreting evidence before passing it onward, making the division of research work part of the system's design.
Key takeaways · AI-distilled
Clavié argues coding agents succeeded first because code has durable, greppable cues (identifiers, file paths) and narrow tasks, while general knowledge work is implicit — a phrase like '30 days' can mean a deadline, grace period, or retention rule depending on context.
In his benchmarkA standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.Full definition → numbers, a badly tuned retrieval tool lands near 60% accuracy — too unreliable to trust — while a well-tuned one reaches the same ceiling as competitors while making 20% fewer tool calls.
Adding a layer of 'searcher agents' that return short memos on top of a well-tuned retrieval tool closes roughly 40% of the remaining gap to human-level performance, per Clavié's figures.
Clavié frames tool and organizational history as one loop — from the Library of Alexandria's catalogue to search engines, and from single polymaths to specialist firms — where better tools create new roles that generate knowledge demanding still better tools.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
benchmark — A standard public test set for comparing AI models — the shared scoreboards behind every "model X beats model Y" claim.
videoVoice AI Where One Wrong Question Breaks the Rules - Shiv Indap, Upstart & Brooke Hopkins, Coval
video78% Wanted This AI Feature. Almost Nobody Used It - Supriya Subudhi, Bumble
videoWhen Running Your Own AI Models Beats Paying for APIs - Steven Latré, imec
videoHow to Decide What Your AI Agent Should Remember - Wojciech Pluta, Oracle
