Concept
Multi-Agent System (MAS)
A network of individual AI agents that each handle part of a task and coordinate to accomplish goals that would be too complex or slow for a single agent.
Added May 18, 2026
A single AI agent, given a long and complex task, must do everything sequentially: plan, research, execute, review, and iterate, one step at a time. Multi-agent systems distribute this work. Different agents specialise in different parts of the problem, run in parallel where possible, check each other's work, and collaborate to produce results that no individual agent could match in the same time.
The key insight is that many complex tasks are naturally decomposable. Writing a research report, for instance, might be broken into: an agent that searches and retrieves relevant sources, an agent that reads and summarises each source, an agent that synthesises the summaries into a coherent draft, and an agent that critiques and edits the draft. These agents can work in parallel on different sources, hand off their outputs to subsequent agents, and the whole pipeline produces a report faster and better than any single agent working alone.
Communication between agents is a central design question. Agents can communicate through shared memory (a common workspace all agents can read and write), through message passing (agents send structured messages to each other), or through a hierarchical orchestrator that routes tasks and collects results. The choice of communication pattern determines how tightly coupled the agents are and how well the system scales.
Division of labour also allows specialisation. Different agents can be given different tools, different context, different personas, or different base models optimised for their specific role. A coding agent might use a model fine-tuned on code. A research agent might be given access to search APIs. A critique agent might be specifically prompted to find flaws. This specialisation typically produces better results than trying to make one agent do everything.
The challenges of multi-agent systems scale with their complexity. Errors in early agents propagate to downstream agents. Agents can contradict each other. Coordinating many parallel agents introduces latency and cost. And ensuring that the overall system behaviour is predictable and aligned with the intended goal becomes harder as the number of agents and interactions grows.
Analogy
A consulting firm assembling a project team. Rather than one consultant doing all the work sequentially, different specialists contribute in parallel: the market analyst researches the industry, the financial modeller builds the projections, the strategist synthesises findings into recommendations, and the editor produces the final deliverable. Each does what they are best at; the project manager coordinates the whole.
Real-world example
AutoGen, Microsoft's multi-agent framework, enables workflows where multiple GPT-4 instances collaborate on software engineering tasks. One agent writes code, another reviews it for bugs, a third runs the code and reports errors back, and a fourth agent synthesises feedback into revisions. The loop continues until the code passes all tests, producing working software through agent collaboration.
Why it matters
Multi-agent systems represent the path from AI that answers questions to AI that completes projects. Tasks that are too long for a single context window, too complex for sequential processing, or better served by specialised tools require orchestrated agent collaboration. As agents become more capable, multi-agent architectures will be how AI tackles the hardest and most valuable real-world work.
In the news
No recent coverage - check back later.
Related concepts