D1 · Agentic Architecture › Cluster 4 · Multi-Agent

4.2 · Orchestration Patterns

The six shapes of multi-agent systems

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.

PatternShapeWhen to use
Hub-and-spokeOne agent, many toolsStart here. Most agent tasks — a single loop calling tools.
Coordinator / sub-agentA coordinator routes to specialists and mergesDistinct specialised roles that a central agent delegates to and combines.
Sequential pipelineStage → stage → stageDependent stages, e.g. research → outline → draft → edit.
Parallel fan-outIndependent sub-agents run at onceGenuinely independent sub-tasks — cut wall-clock by running together.
RouterClassify, then dispatch to the right specialistHeterogeneous inputs where one specialist per class handles the request.
Evaluator–optimizerProducer + critic in a loopQuality matters and can be judged — one agent produces, another critiques/improves.
🔑 Hub-and-spoke is where you start. It's technically a single agent + tools (from 4.1) — the "many tools" case. The other five are true multi-agent escalations.

DIAGRAM Hub-and-spoke & Coordinator/sub-agent

flowchart LR subgraph HS["Hub-and-spoke (start here)"] H["🟢 Agent"] --> T1["🔧 tool A"] H --> T2["🔧 tool B"] H --> T3["🔧 tool C"] end subgraph CS["Coordinator / sub-agent"] C["🟣 Coordinator"] --> S1["🔵 Specialist 1"] C --> S2["🔵 Specialist 2"] S1 --> C S2 --> C end classDef g fill:#d4edda,stroke:#28a745,color:#155724; classDef b fill:#d1ecf1,stroke:#17a2b8,color:#0c5460; classDef p fill:#e2d4f0,stroke:#764ba2,color:#3d1a5b; class H g; class C p; class S1 b; class S2 b;

DIAGRAM Pipeline, Fan-out, Router, Evaluator–optimizer

flowchart TB subgraph SEQ["Sequential pipeline"] direction LR R["Research"] --> O["Outline"] --> D["Draft"] --> E["Edit"] end subgraph FAN["Parallel fan-out"] direction TB CO["Coordinator"] --> A1["Sub-agent 1"] CO --> A2["Sub-agent 2"] CO --> A3["Sub-agent 3"] end subgraph ROU["Router"] direction LR CL{"Classify"} -->|billing| SB["Billing specialist"] CL -->|tech| ST["Tech specialist"] end subgraph EVO["Evaluator–optimizer"] direction LR P["Producer"] --> V{"Critic"} V -->|revise| P V -->|accept| DONE(["✔"]) end classDef n fill:#e2d4f0,stroke:#764ba2,color:#3d1a5b; class R,O,D,E,CO,A1,A2,A3,SB,ST,P n;

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

Trap: Confusing router with parallel fan-out. → A router classifies then dispatches to one specialist; fan-out runs many sub-agents at once. Router = pick one; fan-out = do all.
Trap: Using a sequential pipeline for independent work. → If stages don't depend on each other, a pipeline needlessly serializes them; fan-out cuts wall-clock (see 4.6).
Trap: Forgetting hub-and-spoke is the starting point. → Many "which pattern?" questions want the simplest answer: one agent, many tools.

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.