CONCEPT Six patterns, one spectrum
Once you've chosen to go beyond a single agent, the shape of the system is one of six named patterns. They range from "barely multi-agent" (hub-and-spoke) to fully parallel and iterative. Start at the top of this list and only move down when the task forces it.
| Pattern | Shape | When to use |
|---|---|---|
| Hub-and-spoke | One agent, many tools | Start here. Most agent tasks — a single loop calling tools. |
| Coordinator / sub-agent | A coordinator routes to specialists and merges | Distinct specialised roles that a central agent delegates to and combines. |
| Sequential pipeline | Stage → stage → stage | Dependent stages, e.g. research → outline → draft → edit. |
| Parallel fan-out | Independent sub-agents run at once | Genuinely independent sub-tasks — cut wall-clock by running together. |
| Router | Classify, then dispatch to the right specialist | Heterogeneous inputs where one specialist per class handles the request. |
| Evaluator–optimizer | Producer + critic in a loop | Quality matters and can be judged — one agent produces, another critiques/improves. |
DIAGRAM Hub-and-spoke & Coordinator/sub-agent
DIAGRAM Pipeline, Fan-out, Router, Evaluator–optimizer
TRADE-OFFS Simplicity vs. capability
✅ Simpler patterns (hub-and-spoke, pipeline)
- Fewer agents, lower cost
- Easy to trace and test
- Deterministic-ish control flow
⚠️ Richer patterns (fan-out, evaluator–optimizer)
- Parallelism & quality gains — but more tokens
- Evaluator loops can run many rounds
- More coordination + merging complexity
EXAM TRAPS
PRACTICE
Q1 · Incoming tickets are billing, technical, or account questions; each has a dedicated specialist. Which pattern?
Router. Classify the ticket, then dispatch to exactly one specialist. It's not fan-out — you don't run all three specialists on every ticket.
Q2 · A blog post must go research → outline → draft → edit, each stage depending on the last. Which pattern?
Sequential pipeline. Stages are dependent, so they run in order. Latency is additive, which is inherent to the dependency — you can't parallelize what depends on prior output.
Q3 · You need to raise the quality of generated SQL until it passes a critic. Which pattern?
Evaluator–optimizer. One agent produces the SQL, another critiques it, and the producer revises in a loop until the critic accepts. Best when quality is judgeable.