Single agent or multi-agent?
We'll cover the following
Start with one agent. Split only when it hurts.
Signs it is time to split
- Too many tools (~15 to 20+): the model picks the wrong one and descriptions eat context. A specialist with 5 tools is more accurate.
- Conflicting instructions: "strict compliance checker" and "friendly sales assistant" in one prompt.
- Different permissions: a read-only research agent vs one that can write or spend money. Splitting makes least privilege cleaner.
- Independent subtasks: research three companies in parallel, each sub-agent with a clean context.
- Different models: a cheap fast model to route, a stronger one for hard reasoning.
Patterns
- Router: a small step classifies the request and sends it to the right specialist.
- Orchestrator + workers: a lead agent plans, delegates to sub-agents, and merges the results.
The cost people forget
- More LLM calls: more latency and money.
- Handoffs lose information, since a sub-agent only knows what you pass it.
- Harder to debug; you need tracing across agents.
Check yourselfA support bot has 4 tools. Multi-agent?
No. Four tools is nowhere near the pain point. Multi-agent would only add latency, cost and debugging work.
Summary: say it in the interview
- "I start with a single agent and a good tool set. I split when I see tool-selection errors, context bloat, or a need for isolated permissions. Multi-agent is not free."