Learn AI /AI Agents and Automation /Multi-agent system

Multi-agent system

Intermediate
AI Agents and Automation
Technical readers
Policy officers

Also called: multi agent, agent swarm.

Detail

In five seconds

What it is

A multi-agent system contains multiple software agents that interact, directly or through a control layer, while carrying out a shared or related task.

In plain words

It is like assigning a team instead of one worker: one part searches, another checks, another organises, and their pieces are combined into one result.

An everyday example

Imagine several people preparing dinner. One reads the recipes and sequences the tasks, one prepares ingredients, one uses the oven, and another checks that the dishes will be ready together. They exchange updates because a delay in one task changes what others should do. The arrangement works only when roles and hand-offs are clear: otherwise, two people may prepare the same dish while nobody starts another. A multi-agent system is similar, except the participants are software agents, their tools and permissions may differ, and a controller may assign and combine the work.

Where the story stops working

People in a library usually have broad common sense and can discuss ambiguity in rich ways. Software agents usually rely on narrower instructions, tool outputs and message formats. Also, human team members may truly understand institutional context, while AI agents only generate outputs from patterns and system prompts, so their agreement does not guarantee correctness.

Remember thisA multi-agent system divides related work among interacting agents. The benefit comes from clear task ownership, controlled permissions and reliable hand-offs, not simply from adding more agents.

This term is contested

The broad idea is widely used, but boundaries vary in practice. Some people reserve the term for systems with relatively autonomous agents that communicate with each other, while others use it for any architecture with multiple specialised agent components under one coordinator.

Understand it

Why it matters

For non-specialists, the key issue is not whether the system sounds impressive. It is whether work is being split across components in a way that improves coverage, speed or checking, or whether it simply adds complexity and risk. Multi-agent systems can help with large, messy tasks by assigning different roles, running parts in parallel and comparing outputs. But they can also make responsibility harder to trace. If several agents pass errors to each other, a mistake may look well supported when it is not. In procurement, governance and risk review, you may need to ask: Which agent can access which tools? Who checks final outputs? What triggers human oversight? What logs exist? Those questions affect reliability, accountability, privacy and security more than the label alone.

How a multi-agent system handles one request

Show how one request can be split among specialised agents, coordinated, checked and returned as one result.

01
1. Request enters
A user provides a goal, such as finding prior language on a topic.
Draw one input card on the left labelled 'Request'. Add an arrow to a central box labelled 'Coordinator or planner'.
02
2. Work is divided
The coordinator assigns different sub-tasks to different agents.
From the coordinator, draw three arrows to three boxes: 'Search agent', 'Analysis agent' and 'Policy-rules agent'. Use distinct colours and short role labels.
03
3. Agents use tools and messages
Each agent may call tools or exchange short results with the others.
Place small tool icons under each agent, such as database, document and rules list. Add thin dashed arrows between agents to show message passing.
04
4. Outputs are checked
A checker compares the sub-results for relevance, consistency and missing evidence.
Merge arrows from the three agents into one box labelled 'Checker / verifier'. Add a checklist icon.
05
5. Final result or human review
If the result is strong, the system returns it. If not, it asks for human oversight.
From the checker, branch to two end boxes: 'Answer with sources' and 'Human review needed'. Label the branch condition 'confidence / policy rules'.

A left-to-right flow diagram. One user request enters a coordinator, which sends work to three specialist agents in parallel. Their outputs move to a checker agent, then to a final answer box or a human review box if confidence is low. Read from left to right to see how division of labour and supervision change the result.

Hovering over each agent highlights its role and the tool it uses; clicking shows an example message passed between agents.

Describe this without the diagram

The diagram shows a multi-agent system as a flow from left to right. A request first goes to a coordinator or planner. The coordinator sends different parts of the work to three specialist agents: one searches, one analyses documents and one checks policy rules. These agents can also exchange short messages. Their outputs then move into a checker or verifier that compares the pieces. From there, the process splits into two possible outcomes: either an answer with sources is returned, or the case is sent for human review if confidence is low or rules require oversight. The main lesson is that multiple specialised agents can share one task, but a checking and review stage is needed before acting on the result.

With reduced motion: A static flowchart with the same boxes and arrows, without animation. Parallel work is shown by three side-by-side agent boxes between the coordinator and the checker.

A multi-agent system is a setup where more than one software agent interacts while working on a shared or related task. An agent is a software component that uses its current state, instructions and available tools to select and carry out actions with some autonomy. Instead of one agent handling all decisions and actions, several agents can take different roles.

What goes in can be a user request, a queue of tasks, data from tools or rules about what each agent may do. Then the agents interact directly or through a control layer. One may plan the job, another may retrieve documents, another may compare answers, and another may check quality or ask for human review. They may share messages, partial results or a short record of what has already happened.

What comes out is not just one reply. It can be a drafted report, a ranked list of documents, a set of actions taken through tools, or a recommendation for a human to approve. The benefit is that specialised parts can work in parallel or cross-check each other. The risk is that coordination errors can spread quickly if the design is weak.

This is often confused with orchestration. Orchestration is the control layer that routes tasks and manages order. A multi-agent system is the wider arrangement of multiple interacting agents. It may include orchestration, but it is not only that control layer.

It is also different from a workflow. A workflow defines steps that move work from input to result and may include people, ordinary software or agents. A fixed sequence does not become a multi-agent system merely because it contains several steps or components.

A single chatbot or assistant with many features is not automatically a multi-agent system. The key idea is multiple agents that can select actions within their roles and interact to carry out shared or related work.

How “Preparing dinner together” maps onto the real thing

In the story	In the AI system
person turning the request into search terms	planning or query-generation agent
person searching the catalogue	retrieval agent using tools or databases
person reading relevant pages	analysis agent extracting useful content
person checking topic and date	verification or critic agent

A worked example

An evaluation officer uses an approved multi-agent workspace to check whether three language versions of draft terms of reference remain aligned before the clearance file goes forward.

Starting from: Compare the three versions for missing sections, inconsistent scope, differing dates or deliverables, and terminology that may alter meaning. Produce an issues log with passage references. Do not edit the drafts.

A coordinator agent separates the request into structural, cross-language and consistency checks, then assigns each check to an agent with the relevant instructions and access.
A structure agent compares each version with the office's required headings and reports omissions or sections appearing in only one version.
A terminology agent compares defined terms, responsibilities and deliverables across the versions, marking differences that could change the intended scope.
A critic agent checks each finding against the source passages, rejects unsupported claims and flags cases where language expertise is needed.
The coordinator removes duplicate findings and assembles the verified items into one issues log for the evaluation officer.

Result: The officer receives a consolidated issues log rather than three separate machine reviews. Each item identifies the relevant versions and passages, distinguishes likely inconsistencies from uncertain cases, and leaves all changes and clearance decisions to staff.

In a United Nations workplace

At a duty station working in three languages, an evaluation officer is preparing terms of reference for clearance on Friday. In an approved workspace, a coordinator agent assigns one agent to check required sections, another to compare scope and terminology across the language versions, and a third to challenge inconsistencies. It then returns a single issues log, linked to the relevant passages, for the officer to resolve before sending the file to the front office.

The example shows a centralised multi-agent pattern: agents with different instructions perform separate checks, while a coordinator assigns work and combines their findings. This can make a complex review more systematic without treating any agent's output as the final judgement.

Before you try thisDo not upload unpublished terms of reference to an external tool that has not been approved for them. If a draft contains personal, confidential or procurement-sensitive material, it must remain in a system approved for that material, with appropriate access controls. Automated comparison is not language review or clearance, and every proposed correction requires human verification.

Where you may hear it

During a vendor demonstration to an information technology governance board
In a workflow design session between a programme officer and a technology focal point
At a risk review before an automated research tool is cleared for use
During a procurement unit's technical evaluation of an automation platform

Avoid confusion

Often confused with

Common misconceptions

Common misconception: If several agents work on a task, the answer is automatically more reliable.
Correction: More agents can help with coverage or cross-checking, but only if roles, evidence handling and verification are designed well. Poorly coordinated agents can repeat the same weak assumption, amplify an early error or create false confidence by agreeing with each other.
Common misconception: A multi-agent system is just one chatbot with a fancy name.
Correction: A single chatbot may use tools and still remain one agent handling everything in one loop. A multi-agent system usually has distinct agents with separate roles, prompts, states, permissions or tool access, and those parts exchange work or results.
Common misconception: The orchestrator and the multi-agent system are the same thing.
Correction: The orchestrator is often the traffic controller: it routes tasks, manages order and may collect outputs. The multi-agent system is the larger arrangement that includes multiple agents, their interactions, their tools and often the orchestrator as one component among others.

Quick check

Which description best fits a multi-agent system?

Go deeper

The simple version

It is like assigning a team instead of one worker: one part searches, another checks, another organises, and their pieces are combined into one result.

The technical version

A multi-agent system is a software architecture in which multiple agents, often backed by language models and tool connectors, pursue a shared or related objective through coordination. Agents may have distinct prompts, memory scopes, permissions, APIs, planning strategies or tool sets. Coordination can be centralised, such as a planner assigning sub-tasks, or more distributed, with agents exchanging messages or negotiating task ownership. Typical patterns include planner-worker, researcher-critic, retrieval-analysis-verification and debate-style checking. Inputs may include a user goal, environment state, policies and tool outputs. During execution, agents perform observation and action cycles: they inspect context, call tools, generate intermediate artefacts and hand off results. The system may maintain shared state, per-agent state or both. The main engineering challenges are task decomposition, communication protocol, conflict resolution, error propagation, latency, cost, traceability and human oversight. In practice, a multi-agent system may outperform a single agent on tasks that benefit from specialisation or parallelism, but it can also underperform if coordination overhead outweighs the benefit.

Reviewed external resources

Primary sourceModel Context Protocol (opens in a new tab)Model Context Protocol
This link explains the Model Context Protocol, which suits technical readers who want to understand one way agents and tools can exchange structured context.
Deeperlink checked 2026-07-31
EncyclopaediaMulti-agent system (opens in a new tab)Wikipedia
This link gives a broad introductory overview of the term, useful for readers who want a neutral starting point and basic vocabulary.
Intermediatelink checked 2026-07-31
ArticleBuilding effective agents (opens in a new tab)Anthropic
This link explains how agent-based systems are framed in practice by one AI developer, and suits readers comparing product and design perspectives.
Intermediatelink checked 2026-07-31
VideoVideo explanations: Multi-agent system (opens in a new tab)YouTube search
This search results page helps readers find a range of video explanations of multi-agent systems, and suits people who learn better by hearing and seeing examples.
search results, not one pagelink checked 2026-07-31
DiscussionCommunity discussion: Multi-agent system (opens in a new tab)Reddit search
This search results page helps readers see how practitioners discuss the term informally, and suits readers who want examples, debates and practical questions rather than a single curated explanation.
search results, not one pagelink checked 2026-07-31

Where this came from

Drafted with a language model through an offline pipeline, checked against this site's schema, and every external link above fetched to see whether it answers. Treat it as a starting point and check anything that matters against the primary sources. How this site is built and what that means.

The one thing to remember

A multi-agent system divides related work among interacting agents. The benefit comes from clear task ownership, controlled permissions and reliable hand-offs, not simply from adding more agents.