Anthropic: multi-agent research system

01 · THE PROBLEM

Research is work with an unpredictable path

According to Anthropic, in research it is very hard to predict the necessary steps in advance, and the task requires flexibility to change course and explore lateral connections as the investigation progresses.

The post defines a multi-agent system as several agents (LLMs that autonomously use tools in a loop) working together. In the Research feature, an agent plans the research process and uses tools to create parallel agents that search for information at the same time.

In Anthropic's internal evaluations, these systems stand out on breadth-first queries, which pursue several independent directions at the same time. A system with Claude Opus 4 as the lead and Claude Sonnet 4 subagents outperformed Claude Opus 4 alone by 90.2% on the internal research evaluation.

02 · WHY IT WORKS AND WHAT IT COSTS

Tokens, parallelism, and economic viability

Anthropic states that multi-agent systems work mainly because they help spend enough tokens to solve the problem. On the BrowseComp evaluation, three factors explained 95% of the performance variance: token usage alone explains 80%, and the other two factors are the number of tool calls and the choice of model.

There is a price. In Anthropic's data, agents typically use about 4× more tokens than chat interactions, and multi-agent systems about 15× more. So the task must be valuable enough to pay for the extra performance.

03 · ARCHITECTURE

Orchestrator and subagents in parallel

Research uses the orchestrator-workers pattern: a lead agent coordinates the process and delegates to specialized subagents that operate in parallel.

1. The query reaches the lead agent, which analyzes it and develops a strategy.

2. The lead saves the plan to memory so as not to lose it if the context is truncated when it goes over the limit.

3. The lead creates subagents to explore different aspects of the question at the same time.

4. Each subagent searches independently with the tools, evaluates what it finds, and returns findings to the lead.

5. The lead synthesizes what it received and decides whether more research is needed, which may lead to new subagents or a refined strategy.

6. When there is enough information, the findings go to the citation agent, which identifies the locations of the citations before the result goes back to the user.

Unlike traditional RAG, which fetches a set of chunks most similar to the query to generate the answer, the architecture runs a multi-step search that finds relevant information dynamically, adapts to new discoveries, and analyzes the results.

The lead's plan is saved to memory because, if the context window goes past 200,000 tokens, it is truncated and it is important to retain the plan.

04 · PROMPTS, TOOLS, AND PARALLELISM

How the lead learns to delegate

In the early versions, the agents created 50 subagents for simple queries, scoured the web endlessly for nonexistent sources, and distracted each other with excessive updates. Since each agent is guided by a prompt, prompt engineering was the main lever for improving these behaviors.

The lead needs to know how to delegate: each subagent needs an objective, an output format, guidance on tools and sources, and clear task boundaries. Without detailed descriptions, agents duplicate work, leave gaps, or fail to find the necessary information. Short instructions such as research the semiconductor shortage were often vague enough for subagents to misinterpret the task or run exactly the same searches as other agents.

Since agents struggle to judge the appropriate effort, Anthropic embedded scaling rules in the prompts:

Agent-tool interfaces are as critical as human-computer interfaces, and poor descriptions can send the agent down completely wrong paths. That is why each tool needs a distinct purpose and a clear description. The team created a tool-testing agent that tries to use a faulty tool and rewrites the description to avoid failures, and this process cut task completion time by 40% for agents that used the new description.

To gain speed, two kinds of parallelization were introduced: the lead agent creates 3 to 5 subagents in parallel, instead of in series, and the subagents use 3 or more tools in parallel. These changes cut research time by up to 90% on complex queries.

Agents tend to use queries that are too long and specific, which return few results. The fix was to instruct them to start with short, broad queries and, after evaluating what is available, progressively narrow the focus.

05 · EVALUATION

Evaluate the outcome, not the script

Even with identical starting points, agents can follow completely different valid paths to the goal. That is why flexible methods are needed, ones that judge whether the agents reached the right results with a reasonable process.

Start early and with small samples. A prompt change can raise the success rate from 30% to 80%, and with effects of that size you can notice changes with few test cases. The team started with a set of about 20 queries that represented real usage patterns.

Since research output is hard to evaluate programmatically because it is free text, Anthropic used an LLM judge that scored each output against a rubric: factual accuracy, citation accuracy, completeness, source quality, and tool-use efficiency. A single LLM call with a single prompt, returning scores from 0.0 to 1.0 and a pass or fail verdict, was the most consistent and aligned with human judgment.

Human evaluation finds what automation misses. Testers noticed that the early agents consistently chose SEO-optimized content sites over authoritative but lower-ranked sources, such as academic PDFs and personal blogs, and adding source-quality heuristics to the prompts helped fix this problem.

For agents that change persistent state across multiple turns, Anthropic succeeded by focusing on evaluating the final state instead of turn-by-turn analysis.

06 · PRODUCTION

State, debugging, deploy, and the limits of synchronous execution

Agents are stateful and errors compound. Restarting from the beginning is expensive and frustrating for users, so Anthropic built systems that resume from the point where the agent was when the error occurred, and combines the adaptability of agents with deterministic safeguards, such as retry logic and regular checkpoints.

Agents make dynamic decisions and are non-deterministic between runs, even with identical prompts. Full production tracing made it possible to diagnose why agents were failing and fix problems systematically.

1. Since agents can be at any point in the process during a deploy, traffic is shifted gradually from the old version to the new one, with both running.

2. Regular checkpoints and the ability to resume from the point where the agent was allow continuing after an error, instead of starting over.

3. Production tracing records decision patterns and interaction structures, without monitoring the content of individual conversations.

Today, lead agents run the subagents synchronously, waiting for each set to finish before proceeding. This simplifies coordination, but creates bottlenecks in the flow of information between agents, and the whole system can end up blocked waiting for a single subagent to finish searching. Asynchronous execution would allow additional parallelism, but brings challenges of result coordination, state consistency, and error propagation among the subagents.

The post's appendix describes this practice: subagents call tools to store their work in external systems and return only lightweight references to the coordinator. This avoids information loss in multi-stage processing and reduces the token cost of copying large outputs through the conversation history.

1. In the diagram, the lead delegates a task to a specialized subagent.

2. The subagent calls tools to store its work in external storage.

3. The subagent returns to the lead only a lightweight reference, without copying the large output through the conversation history.