@berryxia: This team's research is somewhat counterintuitive, taking a different approach to LLM research orchestration. An open-source team discovered that directly depriving the smartest orchestrator in a deep research system of search permissions actually allowed the entire system to top the DeepResearch Bench, beating Claude and ChatGPT...

The research from this team is also a bit counterintuitive, and their approach to LLM research scheduling is quite unconventional.

An open-source team discovered that by directly stripping search permissions from the smartest orchestrator agent in a deep research system, the entire system topped the DeepResearch Bench, beating Claude and ChatGPT.

This counterintuitive design makes Onyx the strongest publicly available deep researcher.

It’s called Onyx, completely open-source on GitHub, and you can run it right now.

The story sounds simple, yet it exposes a common flaw in almost all big tech AI Agents.

Traditional deep research systems—including OpenAI’s o1 series, Anthropic’s, and Google’s solutions—equip the orchestrator with a full set of tools: it can search the web, open links, read documents, and write reports end-to-end.

So what happens?

The orchestrator, once it gets search access, can’t resist doing everything itself. It frantically fetches results, skims the surface, never does high-quality task decomposition, and the final report is always superficial.

After observing this fatal bug, Onyx’s team did something no one else dared: they completely removed the search tool from the orchestrator.

Now it can only write task briefs, decompose queries, and evaluate intermediate reports from sub-agents. But it absolutely cannot browse the web, retrieve information, or jump to conclusions prematurely.

This forced the orchestrator to engage in genuine high-level strategic thinking.

The entire architecture maintains only two layers: a pure-strategy orchestrator on top, and up to six independent research agents below.

A three-phase pipeline that’s super clear:

Phase 1: The orchestrator (with no tool permissions) breaks down a complex problem into up to six focused research directions, producing highly self-consistent task briefs.

Phase 2: Distributes tasks to three isolated research agents. Each agent runs up to 8 cycles of “search-read-think”, producing intermediate reports with cited sources. It can also integrate with internal enterprise sources like Confluence, Slack, and 100+ other data sources, with strict document-level access control.

Phase 3: A deterministic step that deduplicates all reports, renumbers them, generates a unified reference map, and outputs the final high-quality report.

Because the orchestrator never touches raw data, it’s immune to the temptation of “stop at the first result”. And with only two layers of information transfer, there’s no distortion from multi-level summarization.

Result: Onyx achieved No.1 on DeepResearch Bench, comprehensively surpassing closed-source Claude and ChatGPT.

Even more impressive: it can seamlessly connect to internal enterprise knowledge bases, something many paid solutions can’t do.

You can try it tonight.

Just go to the Onyx GitHub repo (link below), star it, then follow the README to run the entire system. Use CrewAI for orchestration and Mistral’s Voxtral for voice input/output to replicate a fully open-source top-tier deep researcher.

The entire framework is 100% open source—architecture details, pipeline code, experimental data—all in the repo.

Big Tech is still stuck on “stuffing more tools and more context into the model”, but Onyx, with its deliberately “crippled” orchestrator, shows everyone that the smartest constraints are often the strongest capabilities.

Avi Chawla (@_avichawla): The No. 1 deep researcher beats Claude and ChatGPT with a trick neither uses.

I studied the open-source architecture behind it.

A counterintuitive thing I found is that the orchestrator agent that runs the entire research strategy has no search access.

It can’t query the web.