What is a Multi-agent system?
A multi-agent system splits a job across several AI agents, each with its own role, tools and instructions, coordinated by code or by a lead agent.
In more detail
Splitting work helps when one prompt would grow too long or when steps need different tools and permissions. A reader agent can see documents but cannot write to your ERP, while a posting agent can write but never sees raw email. That separation is a security benefit as much as a quality one. The cost is more moving parts: more calls, more latency, more places for a handoff to lose context. We start with one agent and split only when a test shows a single agent failing at a specific step. Agents that pass messages to each other without a checkpoint compound each other's mistakes, so we put validation between them.
Example
In an accounts payable flow, an intake agent classifies incoming documents, an extraction agent pulls header and line fields, and a matching agent compares them with purchase orders. Plain code checks the totals between each handoff.