Multi-agent patterns
Designs where multiple agent roles collaborate—researcher, critic, executor—via messages or shared state.
What it is
Designs where multiple agent roles collaborate—researcher, critic, executor—via messages or shared state.
Why it matters
Sometimes specialization helps; often it adds failure modes and cost. Know when a single agent + tools is enough.
How it works (plain)
Split roles → define handoffs → shared memory rules → stop conditions → human gates for irreversible acts.
Try it
Design a 2-agent loop for “write a FAQ”: drafter + fact-checker with citation requirement.
Myths
- ⚠️ Myth: More agents always mean more intelligence.
- ✓ Reality: More agents can mean more confusion and spend.
Sources
- Course 10 agents; planning; evaluation-of-agents
- OpenAI Academy: https://academy.openai.com/en ↗