Why is coding easier for AI than legal or medical work?

Code has explicit structures like function names and file paths that act as reliable anchors. Legal or medical texts rely on implicit meaning where the same phrase can mean different things depending on the context.

What is a 'searcher agent'?

A searcher agent is a specialized sub-agent designed to perform specific information retrieval tasks and return a summarized 'memo' instead of raw data, helping the main agent understand context better.

What does Benjamin Clavié mean by the 'tool loop'?

It is the concept that better tools create new professional roles, those roles generate new knowledge, and that new knowledge then demands even more advanced tools.

Why Coding Agents Are Not the Blueprint for Knowledge Work

Explore Benjamin Clavié's argument on why coding agents succeed while general knowledge agents struggle, and how 'searcher agents' can bridge the gap.

We have spent the last two years obsessed with coding agents, assuming that if an AI can fix a bug, it can eventually run a law firm or a marketing agency.

But Benjamin Clavié, from Mixedbread, argues we might be falling into a massive trap by generalizing from the one domain that is actually the easiest to automate.

If we want AI to handle true knowledge work, we can't just build better programmers; we have to design Knowledge Agents.

⚡ TL;DR: Coding agents succeed because code has explicit, searchable cues like identifiers and file paths. General knowledge work is much harder because meaning is implicit. By using searcher agents that return short memos, we can close roughly 40% of the performance gap between AI and human experts.

The Simple Version

Imagine you are looking for a specific person in a highly organized library. Every book has a clear Dewey Decimal number, a title, and an author—you can find exactly what you need with a simple index.

Now, imagine trying to find a specific 'feeling' or a 'vibe' in a giant pile of handwritten diary entries scattered across a room. There are no labels, no categories, and the same word might mean something completely different in every entry.

Coding is like the organized library. It has strict rules, clear labels, and predictable structures that make it easy for an AI to navigate.

Knowledge work—like law, medicine, or strategy—is like the pile of diaries. The meaning isn't written on the surface; it's hidden between the lines, requiring deep context to understand.

🎯 Think of it this way: Coding is searching a structured database where every field is labeled. Knowledge work is searching a conversation where the meaning changes based on the tone and the history of the speakers.

How It Actually Works

The 'Durable Cue' Advantage in Code

Why are tools like GitHub Copilot or Cursor so effective? It is because code is built on durable cues.

When an agent looks at a codebase, it can use grep or semantic search to find specific method definitions, class names, or file paths. These are unambiguous anchors that tell the agent exactly where it is and what a piece of text does.

In code, a function named calculate_tax() almost always does exactly what it says. The instructions are narrow: 'close this ticket' or 'fix this bug.'

The Ambiguity of General Knowledge

In contrast, general knowledge work suffers from extreme ambiguity. Clavié uses the example of an agent searching a legal contract for the phrase "30 days."

In a vacuum, that string is useless. Is it a deadline for a payment? A grace period for a cancellation? A retention rule for a contract? Or did the agent accidentally wander into a section about a medication schedule?

Without the ability to reconstruct the intent of the search, the agent is just matching patterns without understanding purpose.

🎯 Think of it this way: A coding agent is a mechanic looking for a specific part number in a manual. A knowledge agent is a detective looking for a motive in a witness statement.

Technical Deep Dive

The Tool Loop vs. The Organization Loop

Clavié suggests that human progress isn't just about better tools, but a symbiotic relationship between tools and social structures. He identifies two interlocking loops:

The Tool Loop: Better tools (like search engines or bibliographies) lead to better ways to organize information.
The Organization Loop: New roles (like specialists or librarians) emerge to use those tools, which in turn generates new types of knowledge.

For an AI to move beyond a simple chatbot, it must participate in this loop by acting as a specialized role within an organizational structure, rather than just a generic text predictor.

Closing the 'Oracle Gap' with Searcher Agents

The technical challenge is moving from simple Retrieval-Augmented Generation (RAG) to something more sophisticated. Standard RAG often hits a ceiling because it just grabs chunks of text and hopes for the best.

Clavié discusses a more advanced architecture: adding a layer of searcher agents. Instead of the main agent doing all the heavy lifting, these sub-agents perform targeted searches and return short memos rather than raw text blocks.

By summarizing the findings and providing context before the main agent sees the data, this method can close roughly 40% of the remaining gap between AI performance and human expertise. This reduces the cognitive load on the primary model and prevents it from getting lost in irrelevant data.

In terms of efficiency, a well-tuned retrieval system can reach the same accuracy ceiling as its competitors while making 20% fewer tool calls, making it significantly faster and cheaper to run in production environments.

Why This Matters

If we continue to build AI solely based on how well it handles code, we will create a generation of agents that are incredibly smart at logic but completely useless in a boardroom or a courtroom.

The industry has generalized from the one domain—programming—that is the most structured and least representative of human intellect. To reach the next level of utility, we must shift our focus from programming agents to knowledge agents.

This means building systems that don't just search for keywords, but reconstruct intent, manage complex hierarchies of information, and use specialized sub-agents to provide meaningful context.