multi-agent LLM systems

Instead of one model trying to do everything, a multi-agent system splits a job across several LLM agents that each play a role and talk to one another — like a small team where one member researches, another writes, and a third reviews. Each agent gets its own focused instructions and tools, and they pass results back and forth until the shared goal is done.

Common arrangements include a manager (or orchestrator) agent that decomposes the task and delegates pieces to specialist workers, and debate or critic setups where one agent proposes and another challenges. The appeal is focus and separation: a narrowly-scoped agent with a short, clear prompt often behaves more reliably than one giant agent juggling everything at once, and roles can be developed and tested independently.

But more agents means more places to go wrong and more cost: every message is more tokens, errors can ripple between agents, and they can stall in loops of polite agreement or mutual confusion. Multi-agent designs help when a task genuinely splits into distinct roles, and mostly add overhead when it does not.