Anthropic's Research system: citations get their own agent

Why it matters

Most research setups send one agent to search and write in a single pass. Anthropic's system splits the work: the LeadResearcher plans first, saves the plan to Memory, and sends subagents out in parallel, each on its own aspect. Attribution is a separate step after the loop, where the CitationAgent maps claims to source locations and adds citations without changing the text. The trade-off is cost: Anthropic's own data puts multi-agent systems at about 15x the tokens of a chat, worth it only where the value of the task is high enough, and their internal eval showed a 90.2% gain over a single agent. As described in June 2025, the open question is where that 15x bill starts paying for itself.

Roles and reporting lines

Role	Reports to	Job (from the source)
User (human)	-	Submits the query; receives the final results with citations
LeadResearcher	User	Analyzes the query, develops a strategy, saves the plan to Memory, spawns subagents, synthesizes results, decides whether more research is needed
Memory (store)	used by LeadResearcher	Persists the plan, because context beyond 200,000 tokens is truncated
Subagents (1 to n, in parallel)	LeadResearcher	Each searches one aspect independently (web searches, interleaved thinking on tool results) and returns findings
CitationAgent	gets all findings after the loop exits	Processes the documents and the research report and identifies specific locations for citations

Source data: delegation brief per subagent (article, prompt principle 2): objective · output format · guidance on tools and sources · clear task boundaries.

Source data: effort rules in the prompts (article, principle 3):

Query type	Agents	Tool calls
Simple fact-finding	1 agent	3 to 10
Direct comparison	2 to 4 subagents	10 to 15 each
Complex research	more than 10 subagents	clearly divided responsibilities

Routines

Per query: plan → spin up 3 to 5 subagents in parallel (subagents use 3+ tools in parallel) → synthesize → more research or exit
Subagents run synchronously: the lead waits for each set to finish before proceeding (article: "Synchronous execution creates bottlenecks")
After the loop: citation pass on the final report → results with citations returned to the user

The rule worth adopting: spin up 3 to 5 subagents in parallel, each using 3 or more tools, instead of one agent working alone.

Approval gates

Where	Check	Approver	Documented
LeadResearcher	More research needed, or exit the loop	Lead agent	Yes
CitationAgent	Claims matched to source locations before results return	Citation agent	Yes
Human sign-off	-	-	Not described in the article

One gap to close in your own setup: the article describes no human approval step, so a human gate has to be added at the loop exit or before results ship.

Copyable spec

org_chart:
  name: Multi-agent Research system (orchestrator-worker)
  number: 1
  platform: Claude (Anthropic); pattern portable to any stack
  source: https://www.anthropic.com/engineering/multi-agent-research-system
  prompts: https://github.com/anthropics/claude-cookbooks/tree/main/patterns/agents/prompts
  credit: Anthropic Engineering (Hadfield, Zhang, Lien, Scholz, Fox, Ford), 13 Jun 2025
roles:
  - {id: user, title: User, kind: human}
  - {id: lead, title: LeadResearcher, reports_to: user, job: "analyze query, plan, save plan to Memory, spawn subagents, synthesize, decide to continue or exit"}
  - {id: memory, title: Memory, kind: store, job: "persists the plan in case context is truncated (over 200,000 tokens)"}
  - {id: sub, title: Subagent (1..n), reports_to: lead, job: "independent web/tool search on one aspect, returns findings", parallel: true}
  - {id: cite, title: CitationAgent, reports_to: lead, job: "process documents and report, identify locations for citations so claims are attributed; text otherwise unchanged"}
delegation_brief_per_subagent: [objective, output format, tools and sources to use, clear task boundaries]
effort_rules:
  simple_fact: "1 agent, 3-10 tool calls"
  comparison: "2-4 subagents, 10-15 calls each"
  complex: "10+ subagents, clearly divided responsibilities"
routines:
  - per query loop: plan > spawn 3-5 subagents in parallel > synthesize > more research or exit
  - subagents run synchronously (lead waits for each set)
  - after loop: citation pass, then return to user
approval_gates:
  - {where: loop exit, approver: lead agent, documented: true}
  - {where: final report, check: every claim attributed by CitationAgent, documented: true}
  - {where: human sign-off, documented: false}
cost_note: "about 15x the tokens of a chat interaction (Anthropic's data)"

Sources

Anthropic Engineering (Jeremy Hadfield, Barry Zhang, Kenneth Lien, Florian Scholz, Jeremy Fox, Daniel Ford), How we built our multi-agent research system, published 13 Jun 2025. Fetched 15 Oct 2026, 22:23 CEST (HTTP 200).
Anthropic, research prompts in claude-cookbooks: research_lead_agent.md, research_subagent.md, citations_agent.md. Fetched 15 Oct 2026, 22:23 CEST (HTTP 200).

What is confirmed / what is not

Confirmed (article text and repo):

orchestrator-worker pattern; lead agent coordinates, subagents run in parallel
LeadResearcher saves its plan to Memory; context over 200,000 tokens is truncated
lead spins up 3 to 5 subagents in parallel; subagents use 3+ tools in parallel
subagents currently executed synchronously
CitationAgent step after the research loop
"multi-agent systems use about 15× more tokens than chats" (single agents: about 4×)
90.2% improvement over single-agent Claude Opus 4 on Anthropic's internal research eval (Opus 4 lead, Sonnet 4 subagents)
three prompt files public in claude-cookbooks; citations prompt: report text kept identical, only citations added

Not confirmed:

15× and 90.2%: Anthropic's own internal data, not independent benchmarks
the article describes the system as of June 2025; the production system may have changed since
no human approval step described; chart shows agent-level checks only

All org charts · All bot templates

Citations get their own agent.