A multi-agent system is two or more AI agents, each built for a narrower job, coordinating toward one overall task through a defined pattern — passing work to each other in sequence, working the same problem in parallel, or handing off control when the situation changes.
Microsoft’s Semantic Kernel documentation frames the reason for splitting work up this way plainly: “Traditional single-agent systems are limited in their ability to handle complex, multi-faceted tasks. By orchestrating multiple agents, each with specialized skills or roles, we can create systems that are more robust, adaptive, and capable of solving real-world problems collaboratively”. That is a specific vendor framework describing its own design choice, not a Canadian standard — but the underlying shape (specialist agents, a defined way of passing work between them) is the same one any multi-agent build has to choose.
The same documentation names the recurring coordination patterns: a sequential pattern passes one agent’s result to the next in a fixed order; a concurrent pattern broadcasts a task to every agent and collects their results independently; and a handoff pattern, which the same table describes as: “Dynamically passes control between agents based on context or rules” — useful for escalation, where a first agent works a request until it hits a case it isn’t built for and passes it on.
A research task can be split three ways: one agent gathers source documents, a second drafts a summary from what the first found, and a third checks the draft’s claims against the original sources before it reaches a person — each agent has one job, and the sequence hands the output of one directly into the next rather than one agent doing all three.
See also: what is orchestration, what is an AI agent, what orchestration means in AI.
Deciding how many agents a build actually needs, and who owns each one’s job, is a design question custom-ai-solutions covers.
