What is Multi-agent?
Multi-agent systems, told as a story about a meeting where parrots with different roles share the work and one parrot coordinates.
One parrot got everything. The tray overflows and the roles mix.Find beans, write beans and check beans all mix on one tray. The parrot can't tell whether it is researching or reviewing right now.
Pile it all on one tray and the front falls off. And nobody reviews.By the time it writes, the research beans have been pushed off the tray. And a parrot reviewing its own writing misses its own mistakes.
Multi-agent means splitting the job among parrots with different roles, with one parrot coordinating.The research parrot, the writer parrot and the review parrot each get their own note. The host parrot splits the job and gathers the results. Each one is a parrot that plans its own errands.
The host splits and gathers. Parrots talk by note only, and review is separate.A small tray holds only that parrot's beans, so nothing falls off. Passing a result is a single note (a handoff). Jobs that don't wait on each other, like research and writing, run at the same time.
More parrots mean more beans (cost) and more mismatch. Use only as many as you need.Don't use five parrots for a two-parrot job. If the order never changes, an errand checklist is cheaper. How one parrot works: the parrot that plans its own errands; tray size: the tray in front of the parrot.
Multi-agent = splitting a job among several parrots with different roles (research, write, review), with one host parrot that splits and gathers. Each keeps its own tray and they talk by note only. More parrots mean more beans and more mismatch.
An orchestrator coordinates several sub-agents, each with its own prompt and tools. Their context windows are separate and results pass as handoff messages. Independent work runs in parallel and the reviewer is kept apart from the writer. Token cost and coordination errors grow with every agent, so split only as much as needed.
When grown-ups say it
- Multi-agent
- The parrot meeting. Several parrots with different roles share one job.
- Orchestrator
- The host parrot. Splits the job, gathers the results, decides what is next.
- Role assignment
- Research, write, review. Each role has its own note and tools.
- Handoff
- The one note that passes a result. Not the whole tray — just what the next parrot needs.
- Sub-agent
- One parrot at the meeting. Each keeps a small tray. → the parrot that plans its own errands
- Parallel execution
- Running at the same time. Jobs that don't wait on each other run together — faster, but the beans still add up.
- Cost
- Beans per parrot. Five parrots, five times the beans. → the price of beans
- Agent
- How one parrot works. It plans its own errands. → the parrot that plans its own errands