Case 2 — Multi-Agent Research System (Researcher + Writer + Reviewer)

# Case 2 — Multi-Agent Research System > *"Build a system that, given a topic, produces a researched, well-written article. Three agents: Researcher gathers info; Writer drafts; Reviewer critiques. A supervisor coordinates."* --- ## Step 1 — Architecture ``` [Supervisor] / | \ [Research] [Write] [Review] ``` Supervisor decides which agent to call next based on current state. Each specialized agent uses appropriate tools. ## Step 2 — Specialized agents ```python from langgraph.prebuilt import create_react_agent from langchain_core.tools import tool # Research tools @tool def web_search(query: str) -> str: '''Search the web for information on a topic.''' return tavily_search.run(query) @tool def fetch_url(url: str) -> str: '''Fetch and extract text from a URL.''' return fetch_and_parse(url) # Writing tools (none needed — pure LLM) # Review tools (none needed — pure LLM) # Build specialized agents research_llm = ChatAnthropic(model="claude-haiku-4-5") # cheap for research writer_llm = ChatAnthropic(model="claude-sonnet-4-6") # mid-tier for writing reviewer_llm = ChatAnthropic(model="claude-sonnet-4-6") # mid-tier for critique research_agent = create_react_agent( research_llm, tools=[web_search, fetch_url], state_modifier="You are a research assistant. Gather facts, statistics, examples on the topic. Cite sources.", ) writer_agent = create_react_agent( writer_llm, tools=[], state_modifier="You are a writer. Synthesize research into a clear, well-structured 600-word article. Maintain academic tone.", ) reviewer_agent = create_react_agent( reviewer_llm, tools=[], state_modifier='''You are an editor. Review the article for: - Factual accuracy (do claims match the research?) - Structure (clear intro, body, conclusion?) - Readability (clear sentences, good flow?) - Completeness (covers main points?) Output: - "APPROVED" if ready to publish. - "REVISE: " if needs another pass.''', ) ``` ## Step 3 — Supervisor node ```python from typing import Literal class ResearchState(TypedDict): messages: Annotated[list[BaseMessage], operator.add] topic: str research: str article: str review: str next_agent: str revision_count: int def supervisor_node(state: ResearchState) -> dict: decision_prompt = f'''Topic: {state["topic"]} Research done: {"YES" if state.get("research") else "NO"} Article drafted: {"YES" if state.get("article") else "NO"} Last review: {state.get("review", "NONE")} Revision count: {state["revision_count"]} Decide the next step: - "research" if research is needed. - "write" if research is done but article is not. - "review" if article is drafted but not reviewed. - "rewrite" if review says revise (and revision_count dict: result = research_agent.invoke({ "messages": [HumanMessage(f"Research this topic: {state['topic']}")] }) research_summary = result["messages"][-1].content return { "research": research_summary, "messages": [AIMessage(f"[Research complete]: {research_summary[:200]}...")], } def writer_node(state: ResearchState) -> dict: review_feedback = state.get("review", "") write_request = f"Research:\n{state['research']}\n\nTopic: {state['topic']}" if review_feedback and "REVISE" in review_feedback: write_request += f"\n\nPrevious feedback to address:\n{review_feedback}" result = writer_agent.invoke({ "messages": [HumanMessage(write_request)] }) article = result["messages"][-1].content return { "article": article, "messages": [AIMessage(f"[Article drafted, {len(article)} chars]")], } def reviewer_node(state: ResearchState) -> dict: result = reviewer_agent.invoke({ "messages": [HumanMessage(f"Research:\n{state['research']}\n\nArticle to review:\n{state['article']}")] }) review = result["messages"][-1].content return { "review": review, "messages": [AIMessage(f"[Reviewed]: {review[:200]}...")], } ``` ## Step 5 — Routing and graph ```python def route_next(state: ResearchState) -> str: decision = state["next_agent"] if decision == "research": return "research" elif decision == "write": return "write" elif decision == "review": return "review" elif decision == "rewrite": # Increment revision count, route to writer return "write" else: return END builder = StateGraph(ResearchState) builder.add_node("supervisor", supervisor_node) builder.add_node("research", research_node) builder.add_node("write", writer_node) bui