CONCEPT The escalation, one tier up
4.1 lives inside the agent tier from 1.1. Once you've decided you need an agent, the next restraint question is: one agent with a rich toolbox, or several agents coordinating? The architect's default is unchanged — pick the simplest tier that solves the problem. A single agent + tools is simpler, cheaper, and lower-latency; multi-agent is a deliberate escalation you justify, not a starting point.
| Dimension | Single agent + tools | Multi-agent |
|---|---|---|
| Structure | One model loop, many tools (hub-and-spoke) | A coordinator plus specialist sub-agents, each with its own context |
| Cost / latency | Lower — one context, fewer round-trips | Higher — extra agents each cost tokens; sequential ones add latency |
| Predictability | Easier to test and reason about | More moving parts, more failure modes |
| Best fit | Most agent tasks | Genuinely parallel sub-tasks, 5+ step workflows, distinct specialised roles |
CONCEPT The three "go multi-agent" triggers
Multi-agent earns its overhead only when at least one of these clearly holds:
- Genuinely parallel sub-tasks — independent units of work that can run at once to cut wall-clock (see 4.6).
- 5+ step workflow — enough depth that one context window gets bloated or unfocused; decomposition helps.
- Distinct specialised roles — a researcher, a coder, a reviewer, each wanting a different system prompt, toolset, or model.
DIAGRAM The decision tree
independent sub-tasks?"} Q1 -->|Yes| M["🟣 Multi-agent
parallel fan-out"] Q1 -->|No| Q2{"5+ step workflow
bloating one context?"} Q2 -->|Yes| M2["🟣 Multi-agent
pipeline / coordinator"] Q2 -->|No| Q3{"Distinct specialised
roles needed?"} Q3 -->|Yes| M3["🟣 Multi-agent
coordinator + specialists"] Q3 -->|No| S["🟢 Single agent + tools
hub-and-spoke · cheapest"] Q4{"Cost / latency
sensitive?"} -.forces.-> S classDef s fill:#d4edda,stroke:#28a745,color:#155724; classDef m fill:#e2d4f0,stroke:#764ba2,color:#3d1a5b; class S s; class M m; class M2 m; class M3 m;
TRADE-OFFS What multi-agent buys and costs
✅ Multi-agent gains
- Parallelism cuts wall-clock on independent work
- Specialised prompts/tools/models per role
- Context isolation keeps each sub-task focused
- Decomposes 5+ step workflows into digestible units
⚠️ Multi-agent costs
- More tokens — every agent has its own context
- Sequential agents add latency additively
- State must be passed explicitly (see 4.5)
- More failure modes; harder to test
EXAM TRAPS
PRACTICE
Q1 · A support bot needs to look up orders, check inventory, and issue refunds — sequentially, per conversation. Single or multi-agent?
Single agent + tools. The work is sequential and dependent, not parallel, and the three actions are just tools, not distinct specialised roles. One hub-and-spoke agent with three tools is simpler, cheaper, and lower-latency.
Q2 · A research report must gather facts on 8 independent companies, then synthesize. Single or multi-agent?
Multi-agent. The 8 lookups are genuinely independent, so parallel fan-out sub-agents cut wall-clock, and a coordinator synthesizes. This hits the "genuinely parallel sub-tasks" trigger.
Q3 · A stakeholder says "make it as cheap and fast as possible." How does that bias the single-vs-multi call?
Toward single agent + tools. Multi-agent adds tokens and (if sequential) latency. Only override that bias if a hard trigger — true parallelism or genuinely distinct roles — is present.