Stop building coding agents. Build knowledge agents instead.

The proposal. Ben Clavié argues that the AI agent industry has fixated on coding agents while ignoring the broader, more impactful category of knowledge work. He claims that code is a special, narrow form of knowledge with durable cues, while most real-world knowledge work involves ambiguous, contextual information that behaves very differently. This distinction, he believes, has led teams to design agents poorly for non-coding tasks.

Why it matters. Knowledge work—legal research, financial analysis, medical diagnosis—represents the vast majority of the service economy. Clavié argues that designing agents as knowledge workers, borrowing organizational patterns from law firms and hospitals, unlocks far more value than further optimizing coding agents. The stakes are operational: organizations that adopt this pattern may achieve 40% fewer mistakes on complex research tasks, while those that don't will hit a ceiling.

The evidence. Clavié presents two benchmarks. On BrowseComp+, optimized tooling reduced tool calls by 20% and costs by 95% compared to unoptimized baselines, with accuracy reaching 90.2%. On MQA, a multi-agent architecture that decomposes questions and delegates sub-searches to specialized agents reduced the gap between human and agent performance by 40%, from 10 points to 6 points. These are real, measurable improvements from a structured workflow.

The trade-off. The multi-agent orchestration pattern Clavié describes requires co-designing tools and agents together, not bolting search onto an LLM. Teams must invest in query decomposition logic, specialized search primitives (BM25, semantic, multimodal), and agentic harnesses that understand which tool to use when. This is more architectural complexity upfront, with governance and observability challenges that remain unaddressed in the talk.

What's missing. Clavié is the CEO and founder of Mixedbread, the company whose search tool achieved the best benchmark results in his talk. The evidence blends genuine engineering insight with vendor positioning. The talk does not discuss failure modes of multi-agent systems, token cost blowups, latency trade-offs, or how to validate when decomposition actually degrades performance. The benchmarks are also narrow—single-turn, bounded document sets, not the messy, streaming information environments most teams face.

The bottom line. Clavié's core argument—that knowledge work patterns from law and medicine transfer directly to agent design—is compelling and under-explored. The benchmarks show real gains. But the evidence is thin for production scalability. Engineering leaders should explore the multi-agent decomposition pattern on bounded research tasks while staying skeptical of vendor-specific tooling claims until validated outside the demo environment.