Divide and conquer

When one agent struggles with a sprawling task, a natural move is to split the work across multiple specialized agents — a researcher that gathers sources, a writer that drafts, a critic that reviews — coordinated by an orchestrator that routes work between them. Each agent has a focused role, a tailored prompt, and its own tools, much like a small team of specialists instead of one overwhelmed generalist. Frameworks like CrewAI, AutoGen, and LangGraph make wiring this up straightforward.

Common patterns

A few arrangements recur:

• Orchestrator–workers — a manager agent breaks the task into pieces and delegates to workers.
• Pipeline — agents in sequence, each transforming the previous one's output (research → draft → edit).
• Debate / critic — one agent produces, another critiques, improving quality through review.

When it helps — and when it doesn't

Multi-agent setups shine when subtasks are genuinely distinct and benefit from focused roles or separate tools. But they aren't free: every added agent multiplies latency (more LLM calls), cost, and failure modes (more handoffs to go wrong, more chances to drift off-task). A surprising amount of "multi-agent" complexity can be matched by one well-designed agent with good tools.

The honest default: start with a single agent. Make its prompt and tools excellent. Reach for multi-agent only when one clearly can't cope — not because the architecture diagram looks impressive.

More agents means more coordination, and coordination is where things break. Add an agent only when a distinct role earns its keep; otherwise a single sharp agent wins on speed, cost, and reliability.

Try this: Take a task you think needs a "team" of agents and first try to write it as one agent with a clear prompt and two or three good tools. Often it's enough — and when it isn't, you'll know exactly which distinct role to split off first.