Skip to main content

​
Overview

The Team Leader is the main orchestrator in Junkie’s multi-agent system. It receives user messages, delegates tasks to specialized agents, and synthesizes their responses.

​
Initialization

# agent/agent_factory.py:310-327
team = Team(
    name="Hero Team",
    model=model,
    db=db,
    members=agents,
    tools=[BioTools(client=client), CalculatorTools()],
    instructions=get_prompt(),  # System prompt from Phoenix
    num_history_runs=AGENT_HISTORY_RUNS,
    add_datetime_to_context=True,
    timezone_identifier="Asia/Kolkata",
    markdown=True,
    retries=AGENT_RETRIES,
    debug_mode=DEBUG_MODE,
    debug_level=DEBUG_LEVEL,
    enable_user_memories=True,
    memory_manager=memory_manager,
)

​
Configuration

​
Model Selection

The team leader uses a configurable model based on PROVIDER and MODEL_NAME: 
# agent/agent_factory.py:103-124
def create_model(user_id: str):
    """Create a model instance for a specific user."""
    
    if PROVIDER == "groq":
        return OpenAILike(
            id=MODEL_NAME,
            max_tokens=4096,
            temperature=MODEL_TEMPERATURE,
            top_p=MODEL_TOP_P,
            base_url="https://api.groq.com/openai/v1",
            api_key=GROQ_API_KEY,
        )

    # Custom provider
    return OpenAILike(
        id=MODEL_NAME,
        max_tokens=4096,
        temperature=MODEL_TEMPERATURE,
        top_p=MODEL_TOP_P,
        base_url=PROVIDER,
        api_key=CUSTOM_PROVIDER_API_KEY,
    )

​
Tools

The team leader has direct access to:
  • BioTools: User profile and biographical information
  • CalculatorTools: Basic arithmetic operations
Specialized tools (E2B, Firecrawl, MCP servers) are delegated to specific agents.

​
System Instructions

Instructions are fetched from Phoenix (production tag) with fallback to local file: 
# agent/agent_factory.py:126-147
def get_prompt() -> str:
    """Return system prompt content pulled from Phoenix or fallback."""
    prompt_name = "herocomp"

    try:
        fetched = client.prompts.get(prompt_identifier=prompt_name, tag="production")
        if hasattr(fetched, "format"):
            formatted = fetched.format()
        else:
            formatted = fetched
    except Exception as e:
        print("Phoenix prompt fetch error:", e)
        return get_system_prompt()  # Fallback to local

    messages = getattr(formatted, "messages", None)
    if not messages:
        return get_system_prompt()

    content = messages[0].get("content")
    return content or get_system_prompt()

​
Delegation Hierarchy

From system_prompt.md:75-89, the team leader follows this delegation order:
  1. Deep research / real-time web data / complex analysis → pplx-agent
  2. Short code execution / quick runs / math → groq-compound
  3. Complex code / sandboxed execution / file ops → code-agent
  4. Long-context chat history / thread analysis → context-qna-agent
  5. MCP / platform-specific integrations → mcp_agent (if present)

​
Delegation Rules

  • Parallel delegation: Can delegate to multiple agents simultaneously for complex tasks
  • Fallback: If chosen agent fails, attempt next appropriate agent once
  • Transparency: Never reveal internal delegation to users

​
Team Members

The team is composed of specialized agents: 
# agent/agent_factory.py:303-305
if mcp_tools:
    agents = [perplexity_agent, compound_agent, code_agent, context_qna_agent, mcp_agent]
else:
    agents = [perplexity_agent, compound_agent, code_agent, context_qna_agent]

​
Memory Integration

The team leader has memory capabilities enabled: 
enable_user_memories=True,
memory_manager=memory_manager,
This allows the team to:
  • Remember user preferences
  • Recall past conversations
  • Build context over time
See Memory Management for details.

​
Session Management

Sessions are stored in PostgreSQL: 
db = AsyncPostgresDb(
    db_url=async_db_url,
    session_table="agent_sessions",
    memory_table="user_memories",
)
Conversation history is preserved across Discord interactions using num_history_runs.

​
Debugging

Enable detailed logging: 
DEBUG_MODE=true
DEBUG_LEVEL=2
This shows:
  • Agent selection reasoning
  • Tool calls
  • Response synthesis
  • Memory operations

​
Context Awareness

The team leader receives temporal context: 
add_datetime_to_context=True,
timezone_identifier="Asia/Kolkata",
From system_prompt.md:48-73, this enables:
  • Understanding message timestamps
  • Distinguishing past from present
  • Time-sensitive responses
  • Reply context awareness

​
Next Steps

Build docs developers (and LLMs) love

Get started for freeTalk to us