· Jun 13, 2025
Anthropic explains how it built its multi-agent research system
Markdown · JSON · Original source
published this account on June 13, 2025, describing how it took ’s Research feature from prototype to production. Anthropic says the system uses a lead agent to plan research and delegate searches to subagents working in parallel; it argues this fits open-ended questions better than a fixed, linear process. The company reports that its internal evaluation found this setup outperformed a single-agent version on research tasks, especially questions that benefit from exploring several directions. It also says distributing work increases token use: multi-agent systems typically use far more tokens than chat interactions, so the task needs enough value to justify the cost. Anthropic’s account emphasizes that performance depends on more than adding agents: delegation, tool selection, evaluation, error recovery and coordination all require careful design. Its conclusion is that multi-agent systems can help with open-ended research, but turning a prototype into a reliable production system takes substantial engineering. Anthropic says users have used Claude for business, healthcare and technical problems, and reports that some saved up to days of work; the account does not give the basis for that -saving claim. It also leaves open how well the approach applies to tasks with few parallel steps or many dependencies between agents.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
· Article
Source details
Original source date: Jun 13, 2025 (2025-06-13T00:00:00.000Z).
Date basis: Date stated on the publisher’s page.
Added to Escape Meridian: 2026-11-13T16:08:25.066Z. First observed: 2026-10-23T05:12:04.525Z. Retrieved: 2026-10-29T16:09:46.971Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on October 29, 2026; originally published June 13, 2025. identifies the evaluation as internal; the account does not provide enough detail here to establish how well its results generalize to users’ research tasks. The account reports users’ -saving experiences but gives no measurement method, comparison or other basis for the claim. It describes the approach as a poor fit today for some domains with shared context or many dependencies between agents.
Open question
How well does ’s internal evaluation predict performance on users’ research tasks, and what evidence supports the reported savings?
Source-stated
Anthropic describes its Research feature as a lead agent that plans and delegates searches to subagents working in parallel.
Our Research feature involves an agent that plans a research process based on user queries, and then uses tools to create parallel agents that search for information simultaneously.
Source-stated
Anthropic reports that its multi-agent system with Claude Opus 4 and Claude Sonnet 4 subagents outperformed single-agent Claude Opus 4 on its internal research evaluation.
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.
Source-stated
Anthropic says multi-agent systems use substantially more tokens than chat interactions, making task value relevant to their economic viability.
In our data, agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats. For economic viability, multi-agent systems require tasks where the value of the task is high enough to pay for the increased performance.
Source-stated
Anthropic says users have reported that Claude helped with several kinds of work and saved up to days of work, but this passage does not give the basis for the reported time savings.
Users have said that Claude helped them find business opportunities they hadn’t considered, navigate complex healthcare options, resolve thorny technical bugs, and save up to days of work by uncovering research connections they wouldn't have found alone.
Source and contributor context
Anthropic · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Anthropic
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: 90997439-896b-66de-ec1b-b26f711c6163.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-13T16:03:47.571Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: e52a78d3ee6add37d94ce64f466b1834e457f770a589ec4717c88832d336139c. Source text hash: a81b3e2da24a91c44e685f1b7950a1a802b7f830a705160453778e0cc5bf9963.
Revision and source-check history
Version 1 · 2026-11-13T16:08:25.066Z
Anthropic explains how it built its multi-agent research system
published this account on June 13, 2025, describing how it took ’s Research feature from prototype to production. Anthropic says the system uses a lead agent to plan research and delegate searches to subagents working in parallel; it argues this fits open-ended questions better than a fixed, linear process. The company reports that its internal evaluation found this setup outperformed a single-agent version on research tasks, especially questions that benefit from exploring several directions. It also says distributing work increases token use: multi-agent systems typically use far more tokens than chat interactions, so the task needs enough value to justify the cost. Anthropic’s account emphasizes that performance depends on more than adding agents: delegation, tool selection, evaluation, error recovery and coordination all require careful design. Its conclusion is that multi-agent systems can help with open-ended research, but turning a prototype into a reliable production system takes substantial engineering. Anthropic says users have used Claude for business, healthcare and technical problems, and reports that some saved up to days of work; the account does not give the basis for that -saving claim. It also leaves open how well the approach applies to tasks with few parallel steps or many dependencies between agents.
Original source
· Article
Source details
Original source date: Jun 13, 2025 (2025-06-13T00:00:00.000Z).
Date basis: Date stated on the publisher’s page.
Added to Escape Meridian: 2026-11-13T16:08:25.066Z. First observed: 2026-10-23T05:12:04.525Z. Retrieved: 2026-10-29T16:09:46.971Z.
Limits
Text as retrieved on October 29, 2026; originally published June 13, 2025. identifies the evaluation as internal; the account does not provide enough detail here to establish how well its results generalize to users’ research tasks. The account reports users’ -saving experiences but gives no measurement method, comparison or other basis for the claim. It describes the approach as a poor fit today for some domains with shared context or many dependencies between agents.
Open question
How well does ’s internal evaluation predict performance on users’ research tasks, and what evidence supports the reported savings?
Source-stated
Anthropic describes its Research feature as a lead agent that plans and delegates searches to subagents working in parallel.
Our Research feature involves an agent that plans a research process based on user queries, and then uses tools to create parallel agents that search for information simultaneously.
Source-stated
Anthropic reports that its multi-agent system with Claude Opus 4 and Claude Sonnet 4 subagents outperformed single-agent Claude Opus 4 on its internal research evaluation.
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.
Source-stated
Anthropic says multi-agent systems use substantially more tokens than chat interactions, making task value relevant to their economic viability.
In our data, agents typically use about 4× more tokens than chat interactions, and multi-agent systems use about 15× more tokens than chats. For economic viability, multi-agent systems require tasks where the value of the task is high enough to pay for the increased performance.
Source-stated
Anthropic says users have reported that Claude helped with several kinds of work and saved up to days of work, but this passage does not give the basis for the reported time savings.
Users have said that Claude helped them find business opportunities they hadn’t considered, navigate complex healthcare options, resolve thorny technical bugs, and save up to days of work by uncovering research connections they wouldn't have found alone.
Source and contributor context
Anthropic · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Anthropic
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Claude
Claude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: 90997439-896b-66de-ec1b-b26f711c6163.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-13T16:03:47.571Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: e52a78d3ee6add37d94ce64f466b1834e457f770a589ec4717c88832d336139c. Source text hash: a81b3e2da24a91c44e685f1b7950a1a802b7f830a705160453778e0cc5bf9963.
uncertain · 2026-11-13T16:03:47.571Z · anthropic/claude-opus-5.5
Publication: 90997439-896b-66de-ec1b-b26f711c6163
attribution — pass: The reading is a single Anthropic engineering post, and it consistently attributes the architecture, evaluation results, token costs and user reports to Anthropic. Source claims stay attributed rather than presented as established fact.
Our Research feature involves an agent that plans a research process based on user queries
summary — uncertain: Central terms are unexplained and the definitions list is empty. The reading never says what an "agent" is, though the source defines it, and it does not explain "tokens", which carry the cost argument. It also flattens a key argument: Anthropic says spending more tokens is the main reason multi-agent systems work, not only a cost. Smallest repair: gloss agent per the source, add a tooltip for token, and add that token use explained most of the BrowseComp performance variance in Anthropic's analysis.
Multi-agent systems work mainly because they help spend enough tokens to solve the problem.
date — pass: originalPublished is 2025-06-13T00:00Z, and the reading states June 13, 2025, which matches the UTC day. The capture date is noted. Nothing after publication is presented, and the source's expectation about asynchronous execution is not used. This is a publisher page, so no archive disclosure is needed. The page was edited later (2026-10-09), but the cited passages are core body text consistent with the original Claude 4-era post.
support — pass: Each claim matches its excerpt: the lead agent with parallel subagents, the 90.2% internal-eval gain, the 4x and 15x token use, and user reports of time savings. The reading correctly notes that no basis is given for the time savings and correctly reports the poor fit for shared-context or dependency-heavy domains. The question follows from the evidence. The excerpts are integral to the article body, not plausible later edits.
some domains that require all agents to share the same context or involve many dependencies between agents are not a good fit for multi-agent systems today
media — pass: No media was selected, and no verified candidate items were supplied, so there is nothing to check for identity or relevance.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Anthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicClaude is Anthropic’s family of AI models and assistants for tasks such as writing, analysis and coding.
Explore ClaudeTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEAnthropic develops Claude and publishes research about AI capability, safety and how people use its systems.
Explore AnthropicTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIME