What a multi-agent system gives you

The defining characteristics of a multi-agent system:

• Parallel execution: subtasks run simultaneously.
• Delegation: a lead agent decomposes the goal, hands out subtasks, and synthesizes the results.
• Distributed context: each agent works with its own context, usually a subset of the whole.

The strengths are the mirror image of the single agent's weaknesses:

• Parallelization: it can explore many paths at once, cutting the total time.
• Specialization: each agent can be tuned and instructed for a specific job.
• Breadth: it can tackle large, many-sided problems.

And the weaknesses are the mirror image of the single agent's strengths:

• Context sharing is hard: getting the right context to each agent is a real problem (this is the heart of Chapter 3).
• Coordination: agents may duplicate work or make conflicting decisions.
• Cost: it is far more token-intensive. Anthropic measured their multi-agent system using about 15 times the tokens of a standard chat.

The article also mentions that not every multi-agent design uses a lead agent. "Swarm" patterns let agents collaborate peer-to-peer, which mixes traits of both approaches and brings its own challenges.