How We Built a Multi-Agent Research System

Claude's Research feature handles complex questions across the web, Google Workspace, and various integrations. The core is not a single agent working alone from start to finish, but an orchestrator-worker structure where LeadResearcher builds a plan and spins up multiple Subagents in parallel to dig into different directions at the same time.

This approach is strong because research itself is inherently non-linear. Answers can't be found via a predetermined path—strategies must keep changing based on clues discovered during investigation. Each subagent explores a different axis in its own independent context window, then compresses only the important information to hand off to the lead agent, which reduces path dependency and enables broader, deeper exploration.

In Anthropic's internal evaluations, a Claude Opus 4 lead agent paired with Claude Sonnet 4 subagents outperformed a single Claude Opus 4 by 90.2% on research evals. For example, on a task of finding board members of companies in the S&P 500 Information Technology sector, the multi-agent system decomposed the task and found the answer through parallel exploration, while the single agent got stuck in slow sequential search.

A big reason for the performance gain ultimately comes down to using more tokens, and using them more efficiently. According to Anthropic's analysis, 95% of the variance in performance on the BrowseComp eval was explained by three factors, and token usage alone explained 80%. In practice, agents use about 4x more tokens than chat, and multi-agent systems use about 15x more tokens than chat, and upgrading to Claude Sonnet 4 produces a bigger performance gain than doubling the token budget of Claude Sonnet 3.7.

On the other hand, costs grow quickly. So this structure is especially suited to high-value tasks, large-scale information exploration that exceeds a context window, and tasks that require handling multiple complex tools. Conversely, it's less suited to areas that don't parallelize well, such as coding tasks where shared context is essential or where there's heavy dependency between agents.

The structure works as follows:

• When a user submits a question, LeadResearcher devises an approach strategy.
• The plan is saved to Memory so it persists even if the context window exceeds 200,000 tokens and gets truncated.
• LeadResearcher creates multiple Subagents, each assigned a different subtask.
• Subagents repeat web searches and evaluate the results, digging deeper if needed.
• Once enough information is gathered, the results are passed to CitationAgent, which finds citation locations based on the documents and report.
• The final answer is returned along with citations.

There are also clear lessons from prompt engineering. Agents can easily over-act, use needlessly long search queries, keep exploring even when results are already sufficient, or duplicate each other's work. So instead of simple instructions, the lead agent was given explicit subtask, output format, tool/source guidance, and task boundary specifications, along with rules to scale from 1 agent up to 10+ agents depending on query complexity.

Principles that proved especially effective include:

• Think like your agents: Simulated with actual prompts and tools to directly observe failure modes.
• Teach the orchestrator how to delegate: Split subagent roles so they don't overlap.
• Scale effort to query complexity: Used 1 agent for simple fact-finding and scaled up to 10+ for complex research.
• Tool design and selection: Built strict tool descriptions and selection criteria to prevent errors like using web search to find information that only exists on Slack, not the web.
• Let agents improve themselves: Had Claude 4 models propose their own improvements to prompt engineering and tool descriptions.
• Start wide, then narrow down: Began with short, broad searches, then progressively narrowed.
• Guide the thinking process: Used extended thinking and interleaved thinking like a scratchpad to perform planning and re-evaluation.
• Parallel tool calling: The lead agent spins up 3-5 subagents in parallel, and subagents also use 3+ tools simultaneously, cutting complex research time by up to 90%.

Evaluation also had to be designed differently from typical single-agent evaluation. Since a multi-agent system can take a different path each time even with the same input, you can't just look at a fixed correct path. Instead, evaluation needs to check whether the goal is achieved regardless of which valid path is taken, whether tool use and exploration are appropriate, and whether the result quality is sufficient. In the end, this system is a structure that refines parallelism, tool design, prompting, and observability together—both to do research better and to make that research evaluable.