Building Anthropic’s Multi Agent Research System to Outperform Claude Opus 4 By 90%

The following is end-to-end implementation of the blueprint that Anthropic shared for building multi-agent research system, which significantly outperformed single agent workflows of Claude Opus 4.

As we have built more advanced version of this system in-house, we wanted to share the simplified blueprint that you can use as a starting point.

Don’t forget to check out “Closing Thoughts on Implementation” at the end for lessons learned from building advanced version.

Here’s what Anthropic shared in their article, which was a claim to go after.

We found that a multi-agent system with Claude Opus 4 as the lead agent and Claude Sonnet 4 subagents outperformed single-agent Claude Opus 4 by 90.2% on our internal research eval.

And here’s a high-level overview of that multi-agent research system:

The multi-agent architecture in action: user queries flow through a lead agent that creates specialized subagents to search for different aspects in parallel.

We’ll build the simplied version of this system based on the following components:

• Lead Agent that orchestrates research
• Multiple Search Subagents working in parallel
• Memory persistence
• Basic citation tracking
• A REST API built with FastAPI