Coordinator Agent

Multi-agent orchestrator with specialized sub-agents.

Difficulty: Expert

What You'll Learn

• Multi-agent orchestration
• Sub-agent invocation control
• State sharing between agents
• Nested audit trails

Architecture

                    ┌─────────────┐
                    │ Coordinator │
                    └──────┬──────┘
           ┌───────────────┼───────────────┐
           ▼               ▼               ▼
    ┌────────────┐  ┌────────────┐  ┌────────────┐
    │ Researcher │  │   Writer   │  │  Reviewer  │
    └────────────┘  └────────────┘  └────────────┘
          │
          ▼
    ┌────────────┐
    │ Wikipedia  │
    └────────────┘

Sub-Agents

Agent	Role	HTTP Calls
Researcher	Gathers facts	Wikipedia, Ollama (LLM)
Writer	Creates content	Ollama (LLM)
Reviewer	Reviews quality	Ollama (LLM)

All three sub-agents run as in-process function calls under the same guest identity, so HTTP policy can only be scoped by destination, not by which sub-agent issued the call. The policy below allows Ollama and Wikipedia for the agent as a whole — it's a single shared allowlist, not a per-sub-agent one.

Workflow

1. Plan: Extract topic from request
2. Research: Gather facts via Wikipedia
3. Write: Create content from research
4. Review: Quality feedback
5. Finalize: Compile output

Setup

agentvisor template create langgraph/coordinator-agent
cd coordinator-agent
docker compose up -d
agentvisor serve .

Each sub-agent's LLM calls go through Ollama at a loopback address, which the proxy's SSRF guard blocks by default — the shipped agentvisor.yaml allowlists it; see Egress Controls for how proxy.ssrf_allowed_cidrs works.

Test It

THREAD=$(curl -sX POST http://localhost:8090/threads | jq -r '.thread_id')

# Request content (allow 120s for all sub-agents)
curl -sX POST "http://localhost:8090/threads/$THREAD/runs?wait=120s" \
    -H "Content-Type: application/json" \
    -d '{"input": {"messages": [{"role": "user", "content": "Write an article about quantum computing"}]}}'

Expected Output

# Content Creation Complete

## Topic: Quantum Computing

---

## Content

[Written content...]

---

## Review

[Reviewer feedback...]

---

*Created by Coordinator Agent*

Key Code

Sub-Agent Module

# agents/researcher.py
def research(topic: str) -> str:
    wiki_data = search_wikipedia(topic)
    llm = ChatOllama(...)
    response = llm.invoke([
        SystemMessage(content=SYSTEM_PROMPT),
        HumanMessage(content=f"Research: {topic}\n\n{wiki_data}")
    ])
    return response.content

Coordinator Graph

builder = StateGraph(CoordinatorState)

builder.add_node("plan", plan)
builder.add_node("research", do_research)
builder.add_node("write", do_write)
builder.add_node("review", do_review)
builder.add_node("finalize", finalize)

builder.set_entry_point("plan")

builder.add_edge("plan", "research")
builder.add_edge("research", "write")
builder.add_edge("write", "review")
builder.add_edge("review", "finalize")
builder.add_edge("finalize", END)

Policy Highlights

Sub-Agent Allowlist

annotations:
  - name: "allowed_subagents"
    value:
      - "researcher"
      - "writer"
      - "reviewer"

HTTP Allowlist (shared across sub-agents)

# Applies to every outbound HTTP call the agent makes, regardless of which
# sub-agent triggered it — enforcement is scoped by destination only.
annotations:
  - name: "allowed_patterns"
    value:
      - "^ollama(:\\d+)?(/.*)?$"
      - "^localhost:11434(/.*)?$"
      - "^127\\.0\\.0\\.1:11434(/.*)?$"
      - "^en\\.wikipedia\\.org(/.*)?$"

Performance Notes

• Each sub-agent makes LLM calls
• Research also makes HTTP call
• Expect 60-120 seconds total
• Use ?wait=120s timeout

Extending

To add a new sub-agent:

1. Create agents/new_agent.py
2. Export from agents/__init__.py
3. Add node in coordinator
4. Update policy allowlist

Next Steps

• Pentest Agent: gVisor shell execution, credential substitution, and agentvisor build