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 or open URLs.

This looks wrong at first glance. Every other deep research system gives its coordinator far more capability.

For instance:

• OpenAI’s approach trains a single model for many consecutive tool calls. It searches, reads, reasons, and writes the report in a long sequential chain.

↳ The researchers behind the No. 1 system (Onyx) observed that this causes the model to spend cycles on low-value searches instead of maintaining a high-level research strategy.

• Anthropic and Google use an orchestrator-researcher pattern similar to Onyx’s system. The key difference is how aggressively Onyx constrains the orchestrator.

Most orchestrators have access to search and retrieval tools alongside dispatch capabilities. And the moment an orchestrator can search, it will.

So instead of decomposing a query into focused research threads, it starts answering the question itself.

It pulls a few results, skips proper task decomposition, and produces a surface-level report from whatever it found first.

Stripping search from the orchestrator forces it to write self-contained and coherent task briefs for each research agent.

The researchers also kept the architecture only two levels deep. When info passes through multiple agents, each one subtly distorts it through summarization/reinterpretation. Keeping it to two levels prevents this.

These two constraints sit inside a larger three-phase pipeline (the visual below maps this):

→ Phase 1 decomposes the query into up to 6 research directions. No tool access prevents the model from prematurely answering.

→ Phase 2 dispatches 3 isolated research agents. Each runs up to 8 sub-cycles of search, read, and think, to produce an intermediate report with citations.

The agents can also search internal enterprise docs (Confluence, Slack, 100+ connectors) with document-level permissions enforced, unlike proprietary solutions.

→ Phase 3 runs a deterministic step that renumbers and deduplicates to produce a report with a unified citation map.

This pattern has been ranked No. 1 on DeepResearch Bench. The whole implementation is available on GitHub and you can try it yourself.

Here’s the Onyx Repo: https://github.com/onyx-dot-app/onyx…

(don’t forget to star it )

My co-founder wrote a detailed article on building a fully open-source deep researcher using Onyx as the deep research layer, CrewAI for orchestration, and Voxtral by Mistral for voice input/output.

Read it below.

onyx-dot-app/onyx

Source: https://github.com/onyx-dot-app/onyx

Onyx - The Open Source AI Platform

Onyx is the application layer for LLMs - bringing a feature-rich interface that can be easily hosted by anyone. Onyx enables LLMs through advanced capabilities like RAG, web search, code execution, file creation, deep research and more.

Connect your applications with over 50+ indexing based connectors provided out of the box or via MCP.

Deploy with a single command:

curl -fsSL https://onyx.app/install_onyx.sh | bash

⭐ Features

• 🔍 Agentic RAG: Get best in class search and answer quality based on hybrid index + AI Agents for information retrieval
  • Benchmark to release soon!
• 🔬 Deep Research: Get in depth reports with a multi-step research flow.
  • Top of leaderboard as of Feb 2026.
• 🤖 Custom Agents: Build AI Agents with unique instructions, knowledge, and actions.
• 🌍 Web Search: Browse the web to get up to date information.
  • Supports Serper, Google PSE, Brave, SearXNG, and others.
  • Comes with an in house web crawler and support for Firecrawl/Exa.
• 📄 Artifacts: Generate documents, graphics, and other downloadable artifacts.
• ▶️ Actions & MCP: Let Onyx agents interact with external applications, comes with flexible Auth options.
• 💻 Code Execution: Execute code in a sandbox to analyze data, render graphs, or modify files.
• 🎙️ Voice Mode: Chat with Onyx via text-to-speech and speech-to-text.
• 🎨 Image Generation: Generate images based on user prompts.

Onyx supports all major LLM providers, both self-hosted (like Ollama, LiteLLM, vLLM, etc.) and proprietary (like Anthropic, OpenAI, Gemini, etc.).

To learn more - check out our docs!

🚀 Deployment Modes

Onyx supports deployments in Docker, Kubernetes, Helm/Terraform and provides guides for major cloud providers. Detailed deployment guides found here.

Onyx supports two separate deployment options: standard and lite.

Onyx Lite

The Lite mode can be thought of as a lightweight Chat UI. It requires less resources (under 1GB memory) and runs a less complex stack. It is great for users who want to test out Onyx quickly or for teams who are only interested in the Chat UI and Agents functionalities.

Standard Onyx

The complete feature set of Onyx which is recommended for serious users and larger teams. Additional components not included in Lite mode:

• Vector + Keyword index for RAG.
• Background containers to run job queues and workers for syncing knowledge from connectors.
• AI model inference servers to run deep learning models used during indexing and inference.
• Performance optimizations for large scale use via in memory cache (Redis) and blob store (MinIO).

To try Onyx for free without deploying, visit Onyx Cloud.

🏢 Onyx for Enterprise

Onyx is built for teams of all sizes, from individual users to the largest global enterprises:

• 👥 Collaboration: Share chats and agents with other members of your organization.
• 🔐 Single Sign On: SSO via Google OAuth, OIDC, or SAML. Group syncing and user provisioning via SCIM.
• 🛡️ Role Based Access Control: RBAC for sensitive resources like access to agents, actions, etc.
• 📊 Analytics: Usage graphs broken down by teams, LLMs, or agents.
• 🕵️ Query History: Audit usage to ensure safe adoption of AI in your organization.
• 💻 Custom code: Run custom code to remove PII, reject sensitive queries, or to run custom analysis.
• 🎨 Whitelabeling: Customize the look and feel of Onyx with custom naming, icons, banners, and more.

📚 Licensing

There are two editions of Onyx:

• Onyx Community Edition (CE) is available freely under the MIT license and covers all of the core features for Chat, RAG, Agents, and Actions.
• Onyx Enterprise Edition (EE) includes extra features that are primarily useful for larger organizations.

For feature details, check out our website.

👪 Community

Join our open source community on Discord!

💡 Contributing

Looking to contribute? Please check out the Contribution Guide for more details.