Part 5 · Applied AI Systems
Multi-Agent Systems
Several agents can divide up hard work, but teamwork adds new ways to fail.
Key concepts
- An orchestrator divides work among specialized workers.
- Agents can pass work through handoffs, or review one another's output.
- Parallel work can speed up research and broad search.
- Coordination costs tokens, time, and complexity.
- Start with one agent, and add more only when the work truly separates.
The orchestrator-worker pattern divides a task and combines the results.
Formal view and evidence
- Established. fully connected agents have communication pairs, so unstructured coordination grows quadratically.
- Debated. When multi-agent systems outperform a single strong agent with equal resources. Debate can improve some tasks, but agents may also converge on shared errors.
- Open problem. Robust protocols for delegation, shared state, responsibility, and failure containment.
Common misconception
Common belief
"More agents always produce better results."
In practice
Extra agents can add cost and errors when the task does not divide cleanly.
Key takeaways
- Orchestrators divide work among specialized agents.
- Communication overhead grows as more agents interact.
- A single agent is often better for tightly connected tasks.