Agent runtime

Multi-Agent System

An AI architecture where multiple specialized agents collaborate — each with its own scope, tools, and prompt — coordinated by a supervisor or by direct agent-to-agent communication.

Operating principle

Production AI is not a prompt. It is a system of context, tools, permissions, traces, evals, and feedback loops.

What it is

A multi-agent system is an AI system composed of several agents instead of one monolithic agent. Each agent has a narrower scope (one is an expert in retrieval, another in writing, another in code review) and they collaborate through a coordinator or through direct messaging. The hypothesis is that specialization is easier to maintain than generalism.

Why it matters

A single agent trying to do everything has a fatter context window, a longer tool surface, and harder-to-test behavior. Multi-agent designs trade complexity (coordination, message passing, failure modes) for specialization. Whether they're better than a well-engineered single agent is workload-specific.

How it works

Common patterns: supervisor-worker (one agent orchestrates, others execute); peer-to-peer (agents discover and call each other via A2A); pipeline (each agent passes output to the next). Substrates: LangGraph supervisor patterns, Crew AI, OpenAI Swarm, custom orchestration. Cost and latency compound — every additional agent in the loop is another call.

Related resources