Textbook of AI

Multi-agent orchestration is the architectural pattern in which $N \ge 2$ LLM agents collaborate. Each agent has its own context window, system prompt, tool set, and (often) model. Coordination is achieved through messages, shared state, or a supervising agent.

Why multiple agents

1. Specialisation, a "Researcher" agent with web-search tools, a "Coder" agent with Python, a "Critic" agent with no tools but a strong system prompt.
2. Context segmentation, long tasks blow up a single context window; sub-agents work in their own windows and report summaries.
3. Parallelism, independent sub-tasks run concurrently.
4. Adversarial dynamics, debate between agents (Du et al. 2023) improves factuality.
5. Role-play emergent behaviour, Park et al.'s Generative Agents (2023) showed believable simulated societies.

Topologies

Pseudocode (supervisor pattern)

def supervisor(task):
    plan = llm("plan", task)
    results = []
    for step in plan:
        agent = pick_specialist(step)
        results.append(agent.run(step))
    return llm("synthesize", task, results)

Frameworks

• AutoGen (Microsoft, 2023), group chat + tool use.
• CrewAI, role-based, production-focused.
• MetaGPT, software-team simulation, SOP-driven.
• OpenAI Swarm (2024), minimal handoff-based orchestration.
• LangGraph (LangChain, 2024), stateful directed graph of agents.

When not to use multi-agent

Empirical lesson from 2024–2025 production: multi-agent often hurts. Anthropic's 2024 post "Building effective agents" and Cognition's "Don't Build Multi-Agents" both argue that:

• Communication overhead dominates.
• Errors compound across agents.
• A single capable model with good memory management and tool use usually outperforms a pipeline of specialised agents.

The remaining use cases are true parallelism (e.g. searching 100 documents simultaneously) and strict role separation for safety.