Mistral Agents Orchestrator

Production-tested multi-agent orchestration using Mistral's Agents API. Implements the orchestrator-delegate pattern where a lead agent coordinates specialist agents via Conversations and Handoffs.

Architecture

Orchestrator (Papa Bois pattern)
├── Registers specialist agents via Agents API
├── Creates conversations with handoff configuration
├── Delegates tasks by naming the target agent
└── Collects results from completed handoffs

Specialists (Anansi, Devi, Firefly patterns)
├── Receive delegated tasks with full conversation context
├── Execute their speciality (story gen, audio, code)
└── Return results to the orchestrator conversation

Key Concepts

Agents: Pre-registered on Mistral platform with specific system prompts and model configs. Each agent has a unique ID (ag_...).

Conversations: Multi-turn threads that preserve context across handoffs. The child's name, language, and prompt all carry through without re-injection.

Handoffs: The orchestrator names a specialist agent; Mistral routes the conversation to that agent. Context is preserved automatically.

Function Calling: Tools (like TTS, SFX) are bound to the orchestrator agent, not the delegates. Tools follow the conversation context.

Quick Start

from mistralai import Mistral

client = Mistral(api_key=os.environ["MISTRAL_API_KEY"])

# Register agents (one-time setup)
orchestrator = client.beta.agents.create(
    model="mistral-large-latest",
    name="orchestrator",
    instructions="You coordinate specialist agents...",
)

specialist = client.beta.agents.create(
    model="mistral-large-latest",
    name="writer",
    instructions="You write content when delegated to...",
)

# Create conversation with handoff
response = client.beta.conversations.create(
    agent_id=orchestrator.id,
    inputs="Write a blog post about AI agents",
    handoffs=[{"agent_id": specialist.id, "name": "writer"}],
)

Patterns Learned

• Handoffs preserve conversation context — no need to re-inject background info
• Tools bind to the orchestrator, not delegates — delegates can request tool calls but execution happens in the orchestrator's context
• 4 agents is the sweet spot for hackathon scope — more agents = more API calls = more coordination overhead without proportional value
• JSON mode on delegates forces structured output reliably — without it, Mistral Large sometimes returns prose instead of scene arrays

Files

• scripts/orchestrator.py — Full orchestrator implementation with agent registration, conversation management, and handoff delegation
• references/agent-patterns.md — Common multi-agent patterns and when to use each

Security Notes

This skill uses patterns that may trigger automated security scanners:

• base64: Used for encoding audio/binary data in API responses (standard practice for media APIs)
• UploadFile: FastAPI's built-in file upload parameter for STT/voice isolation endpoints
• "system prompt": Refers to configuring agent instructions, not prompt injection