Tools & agents
Multi-Agent Systems
Also known as: agent orchestration, multi-agent AI
Several AI agents working together, often with different roles, to tackle a task that is too big or varied for one agent alone.
A multi-agent system is a team of AI agents rather than a single one. Instead of one AI doing everything, you have several, each with a role, a researcher, a writer, a checker, that coordinate to get a job done, much like a small project team. It can handle bigger, more varied work, but it also adds coordination overhead, so it is not always better than one good agent.
A multi-agent system uses multiple AI agents that collaborate (or are coordinated) to complete a task. Common patterns: - Roles/specialists: different agents for research, coding, review, each with its own tools and prompt. - Orchestrator-worker: a lead agent breaks work into pieces and delegates to workers (orchestration). - Debate/critique: agents check or challenge each other to improve quality.
When it helps, and when it doesn't
Multiple agents can parallelise work and separate concerns, but they add complexity, cost and new failure modes (miscoordination, compounding errors). Practitioner advice: start simple, one agent with good tools, and only add more agents when a task genuinely needs it.
Multi-agent systems compose several LLM-driven agents, potentially with distinct roles, tools, memory and objectives, under a coordination scheme (orchestrator-worker, pipeline, blackboard, or debate). The appeal is decomposition and parallelism: separating concerns, running independent subtasks concurrently, and using critique/verification agents to improve reliability. Interoperability layers (MCP) and orchestration frameworks provide the plumbing.
The hard problems are coordination and reliability, not capability: error compounding across agents and steps, context/state sharing, deadlock and looping, cost/latency blow-up, and evaluation of emergent behaviour. Empirically, added agents help mainly when the task is genuinely parallel or benefits from independent verification; otherwise a single well-scoped agent with good tools is simpler and more robust (Anthropic 2024). Design guidance favours the minimum viable structure, clear interfaces between agents, bounded loops, and human checkpoints for consequential actions.