Case 4 — Hierarchical Multi-Agent Research System

# Case 4 — Hierarchical Multi-Agent Research System > **Executive Briefing:** Build an enterprise competitive intelligence system. A central `Supervisor` coordinates three specialized worker agents: (1) `WebResearcher` (gathers news and financial releases), (2) `PatentAnalyst` (searches patent registries and IP filings), and (3) `BriefWriter` (synthesizes structured executive intelligence memos). The supervisor dynamically delegates tasks, evaluates worker outputs, and enforces revision limits. --- ## 1. Enterprise Scenario & Requirements Monolithic single-agent prompts degrade in accuracy when loaded with 20+ disparate tools. By decomposing capabilities into a supervisor multi-agent architecture: 1. **Specialized Workers:** Author independent worker agents using `create_react_agent`, each equipped strictly with domain-specific tools and persona prompts. 2. **Supervisor Coordination:** Implement the supervisor using `langgraph-supervisor` (`create_supervisor`) with explicit delegation instructions. 3. **Automated Handoffs:** Leverage dynamic handoff tools (`transfer_to_researcher`, `transfer_to_writer`). 4. **Context Window Management:** Configure `output_mode="last_message"` to pass synthesized agent responses rather than flooding the main conversation with dozens of raw tool call tokens. --- ## 2. Explicit Deliverables 1. `workers.py`: Definitions for `research_agent`, `patent_agent`, and `writer_agent`. 2. `supervisor.py`: Compiled hierarchical multi-agent graph with supervisor routing. 3. `test_multi_agent.py`: Test suite verifying: (a) task delegation to appropriate worker based on query intent, (b) multi-worker collaboration on complex queries, and (c) token usage comparison between `last_message` and `full_history`. 4. `sample_intelligence_memo.md`: Executive memo generated by the system on a real technology company. --- ## 3. Starter Code & Scaffold ```python from langgraph_supervisor import create_supervisor from langgraph.prebuilt import create_react_agent from langchain_core.tools import tool from langchain_openai import ChatOpenAI @tool def search_press_releases(company: str) -> str: '''Search recent corporate press releases and news.''' return f"Latest news for {company}: Launched cloud AI platform Q2 2026." @tool def search_patent_database(technology: str) -> str: '''Search USPTO and international patent filings.''' return f"Patent filings for {technology}: 14 new grants in transformer quantization." model = ChatOpenAI(model="gpt-4o") # Create specialized worker agents researcher = create_react_agent(model, [search_press_releases], name="researcher") patent_analyst = create_react_agent(model, [search_patent_database], name="patent_analyst") # Create supervisor workflow = create_supervisor( agents=[researcher, patent_analyst], model=model, prompt=( "You are an executive intelligence director. Assign market news queries to the 'researcher' " "and patent/IP questions to the 'patent_analyst'. Synthesize worker findings into a final briefing." ), output_mode="last_message" ) app = workflow.compile() ``` --- ## 4. 100-Point Grading Rubric | Criterion | Points | Excellent (25 pts) | Acceptable (15 pts) | Needs Work (5 pts) | |---|---|---|---|---| | **Worker Specialization & Tool Scoping** | 25 pts | Workers have distinct personas, isolated toolsets, and clean separation of concerns. | Workers separated but tools overlap heavily. | Single monolithic agent pretending to be multiple roles. | | **Supervisor Handoff Mechanics** | 25 pts | Supervisor correctly identifies appropriate workers and transfers context seamlessly. | Transfers correctly but occasionally loops between workers. | Hardcoded static routing instead of LLM delegation. | | **Context Window Optimization** | 25 pts | Uses `output_mode='last_message'` to suppress raw scratchpads and protect token budgets. | Uses `full_history` causing token usage to balloon. | Manually wipes conversation history, losing context. | | **Synthesis & Memo Quality** | 25 pts | Final output integrates findings from multiple workers into a coherent, cited executive briefing. | Combines text but reads as disconnected paragraphs. | Ignores worker outputs and hallucinates summary. |