Multi Agent Orchestration 101

Build multi-agent systems from scratch in Python 3, then wire them to OpenAI and Claude.

Why I Wrote This Series

I've been building AI-powered systems for a while now — from single LLM wrappers, to RAG pipelines, to autonomous agents. At some point I hit a wall. A single agent was useful, but the moment a problem required parallelism, specialisation, or coordination across different capabilities, I realised I needed something more.

I started reading about multi-agent frameworks. There are good ones — AutoGen, CrewAI, LangGraph. But every time I tried to use them, I found myself cargo-culting abstractions I didn't understand. So I did what I always do when I don't understand something: I built it from scratch.

This series documents that journey. No toy examples, no hand-wavy diagrams. Just real Python, real API calls, and the actual mistakes I made along the way.

---

What You Will Learn

This 5-part series covers everything from first principles to production-grade orchestration:

Part 1: What is a Multi-Agent System? (Pure Python)

• Why a single agent is not enough

• Agent primitives: memory, tools, goals

• Building a minimal agent loop from scratch in Python 3

• Adding inter-agent messaging with a simple queue

• Running two agents that coordinate on a shared task

Part 2: Giving Agents Tools and Memory

• Structured tool definitions (JSON Schema)

• A tool dispatcher that any agent can use

• Short-term vs long-term memory patterns

• Sharing context across agents without a framework

Part 3: Orchestrating Agents with OpenAI

• OpenAI function calling as the tool interface

• Building a supervisor agent that delegates to worker agents

• Parallel vs sequential agent execution

• Handling errors and retries in an orchestrated system

Part 4: Orchestrating Agents with Claude

• Anthropic's tool use API (same concepts, different flavour)

• Claude's extended thinking for planning agents

• Mixing Claude and OpenAI agents in one system

• When to choose Claude over GPT and vice versa

Part 5: Production Patterns

• Structured logging and tracing across agents

• Rate limiting, cost tracking, and token budgets

• Graceful degradation when an agent fails

• Patterns I wish I had known from the start

---

Prerequisites

• Python 3.11+

• Basic understanding of async/await

• An OpenAI or Anthropic API key (free tier works for most examples)

If you have read my LLM API Development 101 series, you already have everything you need.

---

Series Parts

Part	Title	Focus
1	Building Agents from Scratch	Pure Python agent loop, no frameworks
2	Tools and Memory	Tool dispatcher, context sharing
3	OpenAI Multi-Agent Workflow	Function calling, supervisor pattern
4	Claude Multi-Agent Workflow	Tool use API, extended thinking
5	Production Patterns	Observability, cost control, resilience