If we want them to do Knowledge Work, design them as Knowledge Agents — Benjamin Clavié, Mixedbread

An agent searching a contract for "30 days" has no way to know whether it has found a deadline, a grace period, or a retention rule, or wandered into a medication schedule entirely. Benjamin Clavié uses that to explain why coding agents turned out to be the easy case. Code carries durable cues: identifiers, file paths, method definitions, things you can literally grep for. It also comes with narrow instructions, because nobody asks a coding agent to invent a paradigm, they ask it to close a ticket. Knowledge work outside code is the opposite on every axis. The meaning is implicit, the same string means different things in different domains, and the search has to start from an intent the agent must reconstruct on its own. His argument is that the industry generalized from the one domain that least resembles the rest. The good news is that humans solved this already, repeatedly. Clavié traces a tool loop running from the first catalogue of the Library of Alexandria through bibliographies to search engines, and an organization loop running from the single gifted polymath through monasteries and universities to the modern specialist firm. They are really one loop: better tools create new roles, new roles generate knowledge, new knowledge demands better tools. His benchmark numbers make the practical version concrete. A badly tuned retrieval tool lands near 60 percent accuracy, which nobody would trust enough to keep using. A well tuned one reaches the same ceiling as its rivals while making 20 percent fewer tool calls, and adding a layer of searcher agents that return short memos closes roughly 40 percent of the remaining gap to a human. Speaker info: - https://x.com/bclavie - https://mixedbread.com - https://ben.clavie.eu Timestamps: 0:00 - Coding agents are a special case of knowledge work 3:48 - Why code is unusually easy to search 5:05 - Thirty days, and when one clue means four things 5:55 - The tool loop and the organization loop 8:06 - Tooling is never a neutral add on 9:22 - A leaderboard, and two different BM25s 11:14 - How a law firm breaks down a case 12:42 - When even a human hits the tool ceiling 13:31 - Searcher agents, memos, and the oracle gap 15:13 - Design for knowledge work, not for programming