3.7 — When NOT to use multi-agents ⚖️

Multi-agents are impressive, but they’re not a switch to flip by reflex.

🟢 In plain words — adding agents is not “more brains = better.” Each agent costs tokens (the units of text that AI providers charge for — so money and time), and the more there are, the more they can get in each other’s way. The rule: start simple, only add complexity if the task truly demands it.

• It’s expensive. A single agent already consumes ~4× more tokens than a simple chat; a multi-agent system, ~15× (sometimes far more). More agents = more calls = more cost and latency.
• It can degrade quality. On sequential tasks, coordination creates errors. Anthropic measured a performance drop of ~35% when a model is put in a multi-agent setup on a planning task (measured on one specific case — an order of magnitude, not a universal law; same for the ~4×/~15× above).
• The pro rule: start with a single well-designed agent. Move to multi-agents only when the task is broad and parallelizable (e.g., researching 10 topics at once), not when it’s a simple sequence of steps.

Tip

🎯 Key takeaway: “a single agent done well” beats “five poorly coordinated agents” in most real cases. Multi-agents are a specialist tool, not a reflex.

❓ Check your understanding — “the more agents you add, the better the result.” True or false?

See the answer

False. Multi-agents cost much more (~15× the tokens of a simple chat) and can degrade sequential tasks (Anthropic measured −35% on a planning task). Reserve them for broad, parallelizable tasks.

✅ In short

• Multi-agents are expensive (~15× the tokens) and slower.
• On sequential tasks, they can degrade quality (−35% measured).
• Pro rule: start with a single agent; go multi-agent only if the task is broad and parallelizable.