Multi-Agent System
An architecture where multiple AI agents work in parallel or in sequence, each handling a specialised subtask, coordinated by an orchestrator.
A multi-agent system uses more than one AI model instance to complete a task, with each agent specialised for a specific role. Rather than one model doing everything serially, work is divided and parallelised.
Common patterns: - Orchestrator + workers: a planning agent breaks a task into subtasks and delegates to specialised worker agents - Critic/review loop: one agent generates output, a second agent checks it for errors or quality - Parallel research: multiple agents research different subtopics simultaneously, results merged by orchestrator - Pipeline chains: output from agent A becomes input to agent B in a defined sequence
When to use multi-agent: - Tasks too long for a single context window - Tasks with independent parallel workstreams - Tasks requiring cross-checking (one agent writes, another validates) - Specialised agents with different tools or personas
The cost: Coordination overhead — more agents means more API calls, more opportunity for errors to compound, and harder debugging.
Example
A research tool uses three agents: one searches the web, one reads and summarises papers, one compiles a structured report. Total time is shorter than a single serial agent because the first two run in parallel.
Related terms
Agentic AI
An AI system that can break down goals into steps, use tools, and iterate until a goal is achieved.
Agent Orchestration
Coordinating multiple AI models, tools, or agents to work together on a complex goal.
Tool Use / Function Calling
Enabling an AI to call external functions or APIs to perform tasks beyond text generation.