Knowledge agents need a different shape than coding agents. The harder problem is not search.

In an AI Engineer conference talk posted this month, Benjamin Clavié, a researcher at the search company Mixedbread and the author of the open-source retrieval project RAGatouille, argued that knowledge work has been badly underserved by agent design. The field, he said, keeps building coding agents: search, run a tool, check the output, repeat, against a well-scoped ticket. Knowledge work, in law, medicine, finance, and research, is different. The task is not pre-specified. What matters is contextual. He argued these agents need to decompose a problem and orchestrate sub-searches the way a senior professional delegates to associates, not flatten everything into one search-then-answer loop. We checked the talk against external sources, both supportive and competing. The architecture claim holds up well. Several of the talk's own headline numbers do not appear anywhere outside the talk itself. And the sharper finding, from a paper the talk never mentions, is that getting the search right is not the same problem as getting the answer trusted.

#The claim that holds up

Clavié's framing is not merely a talk's own spin. Real, independent evidence backs the decomposition argument on three separate fronts.

The first is MADQA, a multimodal agentic document QA benchmark released by Hugging Face and Snowflake in collaboration with Oxford, UNC-Chapel Hill, and other academic and industrial partners. The talk's auto-generated captions render the benchmark as "MQA," which is a caption-transcription artifact for MADQA. On that benchmark, an agent that iteratively decomposes and searches, Gemini 3 Pro Agent, reaches 82.2% accuracy, ahead of a static retrieval-augmented baseline at 78.6%. The paper's own language is blunt: there is "a nearly 20% gap that neither humans nor LLMs can close," with humans needing one or two search rounds to reach an answer that takes agents up to nine rounds, roughly five times the effort, to match. Decomposition wins over flat retrieval. It is also expensive, and the ceiling is real.

The second is the public leaderboard for BrowseComp-Plus, a peer-reviewed retrieval benchmark (ACL 2026 Main) that Clavié's own employer has a real, submitted, and checkable entry on. The top of that 84-row board is not held by single-agent default retrieval. It is held by multi-agent, hybrid scaffolds: AI21's eight-agent setup at 95.18% accuracy, a multi-agent swarm from Sail Research at 90.72%. Mixedbread's own submission scores 90.48% accuracy and 83.03% recall at an average of 11.53 search calls per question, which is not the lowest call count on the whole board (35 of 84 rows use fewer), but is the third-lowest among every row that scores 80% accuracy or higher. That is a real, independently checkable data point, and it is close in shape to what the talk describes, even if the talk's own framing, roughly 90% accuracy with 20% fewer tool calls, is not identical to the board.

The third is Harvey, a legal-AI company building and benchmarking agents specifically for legal knowledge work, distinct in architecture and evaluation from any coding-agent product. Its existence corroborates the talk's underlying diagnosis: knowledge work is a large, funded, actively benchmarked category worth designing agents for on its own terms, independent of what Clavié or his employer has to gain from that answer being true.

#Where the specific numbers stop holding up

Two of the talk's own headline figures could not be found published anywhere outside the talk. The claim that an agent scores 88.9 against a human's 99.4 on the MADQA-adjacent benchmark, and that Mixedbread's own multimodal intervention improves accuracy by 3.5 points while cutting what the talk calls the oracle gap by 40%, do not appear on Snowflake's blog, on Mixedbread's site, or in the arXiv paper. They read as an internal, unpublished ablation, presented as slide content. That does not make them false. It makes them a vendor's own claim about its own product, stated in a talk, not an externally verifiable statistic, and worth treating as exactly that if you repeat it.

The bigger discrepancy is the talk's progression for BM25 on the BrowseComp-Plus benchmark: a poorly tuned baseline at 60%, a tuned version at 70 to 80%, and a best hybrid harness at 98%. None of those numbers reconcile with the public leaderboard. Plain default-scaffold BM25 entries on that same board score far below 60% (one row, GLM-4.7 with default BM25, scores 33.25%), and the highest score on the entire 84-row board, as of this check, is 95.18%, not 98%. The most likely explanation is that the talk is describing an internal, more favorably configured experiment rather than reading numbers off the public board. The directional point, that a well-optimized hybrid retrieval setup beats an unoptimized one by a wide margin, is consistent with the board's actual shape. The specific numbers quoted in the talk are not reproducible from anything publicly indexed.

None of this undoes the architecture argument. It does mean the talk's evidentiary weight rests more on real, external, independently checkable sources than on the internal figures used to punch up the story, and a reader repeating the internal numbers as settled fact would be repeating a vendor's slide, not a verified result.

#The gap the talk never mentions

The most useful finding in this research did not come from the talk at all. It came from L-MARS, a legal multi-agent system paper (accepted at the AI4Law workshop, ICML 2026) that runs exactly the kind of decomposed, orchestrated knowledge-work agent Clavié argues for, then audits it against a stratified 100-question bar exam set.

The result complicates the talk's whole framing. On that audit, retrieval barely moves accuracy. What moves the numbers is a multi-turn judge loop, a verification pattern layered on top of the search: it lifts strict citation F1 from 0.13 to 0.25 and cuts the no-citation rate from 34% to 13%. A companion case study is sharper still. A single-agent web-search baseline collapses to a citation F1 of 0.22 under external audit, even when its search is working. Retrieve-then-draft pipelines saturate near 0.75 on the same measure.

Read those two numbers side by side and the point is unavoidable: an agent can find the right document and still produce an answer nobody should trust without independently re-checking every citation. Search accuracy and citation faithfulness are different axes. A knowledge agent that gets better at decomposing search still has to clear a second bar the talk never names, which is whether its final claims actually trace back to what it retrieved.

#What an operator does with this

If you are evaluating or building a knowledge-work agent, decomposition and orchestration are the right architectural bet, and the evidence for that claim does not depend on any single vendor's slide. Three independent sources (an academic multimodal benchmark, a public retrieval leaderboard, and a real legal-AI company's product line) converge on it.

But a benchmark score for retrieval accuracy is not a substitute for auditing what the agent's final answer actually cites. L-MARS's own numbers show a system whose search works can still hand back an answer with a strict citation F1 of 0.22. If a knowledge agent is going into a workflow where someone signs their name to the output, the retrieval architecture is only half the design problem. The other half is a step that checks the agent's own citations against the documents it actually pulled, before the answer reaches a human who is trusting it. Decomposing the search without auditing the answer is a half-measure dressed up as a full one, and it is the exact place a knowledge agent that looks good on a leaderboard can still fail the person reading its output.