Multi-Agent Research System

Multi-Agent Research System

graph LR
    subgraph "Multi-agent research system"
        direction LR

        subgraph " "
          direction LR
          Citations["Citations
subagent"]
          subgraph "Search subagents"
              direction TB
              Search1["Search subagent"]
              Search2["Search subagent"]
              Search3["Search subagent"]
          end
        end

        subgraph " "
            direction LR
            LeadAgent["Lead agent
(orchestrator)

Tools: search tools + MCP tools
+ memory + run_subagent +
complete_task"]
            Memory["Memory"]
        end

    end

    subgraph "Claude.ai chat"
        UserChat("What's on your mind tonight?

what are all the companies in the
united states working on AI agents in 2025?
make a list of at least 100. for each
company, include the name, website,
product, description of what they do,
type of agents they build, and their
vertical/industry.")
    end

    %% Connections
    UserChat -- "User request" --> LeadAgent
    LeadAgent -- "Final report" --> UserChat
    LeadAgent --> Citations
    LeadAgent --> Memory
    LeadAgent --> Search1
    LeadAgent --> Search2
    LeadAgent --> Search3

    Search1 --> Search1
    Search2 --> Search2
    Search3 --> Search3

    %% Styling
    style Citations fill:#333,stroke:#666,stroke-width:2px,color:#fff
    style LeadAgent fill:#333,stroke:#666,stroke-width:2px,color:#fff
    style Memory fill:#333,stroke:#666,stroke-width:2px,color:#fff
    style Search1 fill:#333,stroke:#666,stroke-width:2px,color:#fff
    style Search2 fill:#333,stroke:#666,stroke-width:2px,color:#fff
    style Search3 fill:#333,stroke:#666,stroke-width:2px,color:#fff
    style UserChat fill:#333,stroke:#666,stroke-width:2px,color:#fff

Multi-Agent System Process Diagram

%%{init: {'theme': 'dark'}}%%
sequenceDiagram
    participant User
    participant System
    participant LeadResearcher
    participant Subagent1
    participant Subagent2
    participant Memory
    participant CitationAgent

    User->>System: send user query
    System->>LeadResearcher: Create Lead Researcher

    loop Iterative Research Process
        LeadResearcher->>LeadResearcher: think (plan approach)
        LeadResearcher->>Memory: save plan
        Memory-->>LeadResearcher: retrieve context

        LeadResearcher->>Subagent1: create subagent for aspect A
        LeadResearcher->>Subagent2: create subagent for aspect B

        par 
            Subagent1->>Subagent1: web_search
            Subagent1->>Subagent1: think (evaluate)
            Subagent1-->>LeadResearcher: complete_task
        and 
            Subagent2->>Subagent2: web_search
            Subagent2->>Subagent2: think (evaluate)
            Subagent2-->>LeadResearcher: complete_task
        end

        LeadResearcher->>LeadResearcher: think (synthesize results)

        alt More research needed?
            Note over LeadResearcher: Continue loop
        else 
            Note over LeadResearcher: Exit loop
        end
    end

    LeadResearcher-->>System: complete_task (research result)
    Note over System, CitationAgent: Process documents + research report to identify locations for citations
    System->>CitationAgent: Return report with citations inserted
    System->>System: Persist results
    System-->>User: Return research results with citations

The Power vs. Cost Trade-off

%%{init: {'theme': 'dark'}}%%
graph TB
    subgraph "The Power vs. Cost Trade-off"
        direction TB

        subgraph "Performance on Complex Queries"
            direction TB

            subgraph " "
                direction LR
                SingleAgent["Single-Agent System"]
                MultiAgent["Multi-Agent System"]
            end

            subgraph " "
                direction LR
                Result1((X))
                Result2((✓))
            end

            SingleAgent -- "Fails to find answer" --> Result1;
            MultiAgent -- "+90.2% performance" --> Result2;
        end

        subgraph "Relative Token Cost"
            direction LR
            A["Standard Chat
(Baseline: 1x)"]
            B["Single-Agent System
(Approx. 4x)"]
            C["Multi-Agent System
(Approx. 15x)"]

            style A height:50px,fill:#87CEEB, color:#000
            style B height:80px,fill:#FFA500, color:#000
            style C height:120px,fill:#FF6347, color:#000
        end
    end

    %% Styling for nodes and links
    style SingleAgent fill:#f9f,stroke:#999,stroke-width:1px, color:#000
    style MultiAgent fill:#9f9,stroke:#999,stroke-width:1px, color:#000
    linkStyle 0 stroke:red,stroke-width:2px
    linkStyle 1 stroke:green,stroke-width:2px

The AI Agent Evaluation Pyramid

%%{init: {'theme': 'dark'}}%%
graph TD
    subgraph "Evaluation Strategy"
        A["Level 1: Small-Sample Testing (Immediate)
Finds major issues quickly

- Use ~20 representative queries
- Catches large-impact failures
- Enables rapid iteration"];
        B["Level 2: LLM-as-Judge (Scaled)
Grades outputs against a rubric

- Factual & citation accuracy
- Completeness of the answer
- Tool efficiency & source quality"];
        C["Level 3: Human Evaluation
Catches nuance and edge cases

- Identifies source quality bias
- Finds subtle failures
- Uncovers unexpected behaviors"];

        A --> B;
        B --> C;
    end

    %% Styling for dark theme
    style A fill:#90ee90,stroke:#aaa,stroke-width:1px,color:#000
    style B fill:#66CDAA,stroke:#aaa,stroke-width:1px,color:#000
    style C fill:#3CB371,stroke:#aaa,stroke-width:1px,color:#000

Orchestrator Agent's Prompting Heuristics

%%{init: {'theme': 'dark'}}%%
graph TD
    A["1.Plan & Delegate"] --> B["2.Allocate Resources"];
    B --> C["3.Select Tools"];
    C --> D["4.Execute Search Strategy"];
    D --> E{"5.Self-Correct
Evaluate results"};

    E -- "No, task complete" --> F((Finish));
    E -- "Yes, refine strategy" --> C;

    %% Styling for the nodes
    style A fill:#333,stroke:#888,stroke-width:2px,color:#fff
    style B fill:#333,stroke:#888,stroke-width:2px,color:#fff
    style C fill:#333,stroke:#888,stroke-width:2px,color:#fff
    style D fill:#333,stroke:#888,stroke-width:2px,color:#fff
    style E fill:#5c3c3c,stroke:#c88,stroke-width:2px,color:#fff
    style F fill:#2a523a,stroke:#7c7,stroke-width:2px,color:#fff

Sources:

• How we built our multi-agent research system