multiagent rendszer (MAS)
in English: multi-agent system
A system in which multiple agents divide tasks among themselves and typically solve problems collaboratively, building on each other's work.
In plain terms
Imagine a flat renovation. No single handyman does everything: there is an electrician, a tiler, and a painter, plus a contractor who organizes the work and takes over at the end. A multi-agent system does the same in software: multiple AI agents work together, each handling its own subtask.
It is good because large, complex tasks become more manageable this way. One agent gathers material, another writes the text, and a third checks it. Each focuses only on its own task, so it is less likely to get lost in details, and some subtasks can run in parallel.
Compared to a single agent, the difference is that here work distribution and coordination between agents are also part of the system. This means more organization, more computation, and more potential for errors, so it is not worth it for every task.
In the broader professional sense, it is not just about language models: robots, simulations, or trading programs can also form multi-agent systems, where agents may even compete or negotiate with each other.
How it works
- In the most common architecture today, a task reaches a controlling agent, which plans how it can be broken down and who should perform each part. (There are also systems consisting of peer agents without a central controller.)
- The coordinator assigns subtasks to specialized agents; each receives its own instructions, its own tools (such as a search engine, database, or code runner), and its own working memory.
- Agents work sequentially or in parallel and pass their results to each other via messages or shared notes.
- There is often a separate checking agent that reviews the work of others, looks for errors, and if necessary, sends it back for correction.
- At the end, the coordinator merges the partial results into a single answer and passes it to you.
An everyday example
You ask an AI assistant to prepare a thorough comparison of five family cars. For some research functions, multiple agents search for each model in the background, another organizes the data into a table, and a third verifies the sources. You only see that the 'deep research' works for a few minutes, then you receive a finished report.
Who is affected
A common misconception
Limits and risks
- It is more expensive and often slower, because each agent consumes separate computing capacity, and coordination also requires resources.
- Errors cascade: if one agent makes a mistake, others can build on it, and the final result can sound confident but be wrong.
- It is harder to trace what each agent did and why, making debugging and liability determination more complex.
- Multiple agents access more tools and data, increasing the attack surface, so it is important that humans approve sensitive steps.
Similar, but not the same
- AI-agent – An AI agent is a single independently acting program, while a multi-agent system is the organized collaboration of multiple such agents.
- agent orchestration – Orchestration is the method of coordinating the work of agents, while a multi-agent system is the coordinated agent team itself.
- chatbot – A chatbot is the conversational interface through which the user asks questions; behind it, a single model, an agent, or even a full multi-agent system may operate.
Related concepts
agent-based AI, artificial intelligence, elosztott rendszer
orchestrator agent, alagent, agentraj
AI-agent, agent orchestration, coding agent, tool use, MCP (Model Context Protocol), human oversight
A brief history
The concept grew out of distributed artificial intelligence research in the 1980s, long before today's language models. It gained new momentum from 2023, when frameworks for agents built on large language models and conversing with each other appeared, such as Microsoft researchers' AutoGen system.
Where things stand
At the end of 2026, it is already common in research and coding AI tools for several agents to work together in the background, and this is spreading in corporate processes too. However, because of reliability, cost and transparency concerns, most organisations still use them for well-defined tasks under human supervision.
In the news
- MAS Future of Finance Institute starts operations in Singapore to speed up AI adoptionNov 3, 2026 · Money
- Bank of Japan, MAS and China's central bank focus on AI risks and strategyOct 20, 2026 · Money
- Singapore's MAS issues risk management guidelines on banks' use of AIOct 7, 2026 · Money
- Singapore's MAS reviews deepfake scam defences as cross-bank AI shield is testedSep 22, 2026 · Money
- MAS tests cross-bank AI models to detect scams fasterSep 14, 2026 · Money
- MAS and industry publish SAFR framework for runtime safeguards on AI agents in financeAug 31, 2026 · Money
- ECB researchers warn an AI-rally correction could hit the euro area hardAug 17, 2026 · Money
- Singapore: MAS SAFR runtime governance standard for agentic AI and a new AI taskforce for cyber resilienceAug 17, 2026 · Money
Also known as
multiagent rendszer · multi-agent rendszer · multi-agent system · multi-agent system · multi-agent system · multiagent system · MAS · agentcsapat
Sources
This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.