Multi-agent systems
Multi-agent systems with rolesand supervision.
A multi-agent system splits a process between several specialised AI agents, for example one for research, one for drafting and one for checking. Each does a single job and has access only to what it needs, and a person approves the final result. We only propose it when a single agent is no longer sufficient.
A single agent with too many responsibilities becomes hard to control
As an agent is given more tools and instructions, it becomes harder to test and to understand why it made a particular decision. On top of that, giving it access to everything increases the risk. Splitting the work by role solves this, but only if the process justifies it.
- Overly long instructions and too many tools reduce the consistency of a single agent
- A single agent with broad access contradicts the principle of least privilege
- Without an independent checking stage, errors get passed on
- Without clear roles, it becomes hard to tell which part of the system made the mistake
Who it is for
- Companies that already have an AI agent in production and are reaching its limits on more complex processes
- Processes with distinct stages that need different tools and permissions (for example, public research and sensitive internal data)
- Teams that need a checking stage separate from the content production stage
- Organisations that want granular control: who can read, who can write, who can only propose
What a multi-agent AI system does
- 01
Splits the process by role
A coordinator agent receives the request, breaks it into tasks and sends them to the specialised agents.
- 02
Separates data access
Each agent uses only the tools and data of its role, for example read-only access or public sources only.
- 03
Checks another agent's work
A checking agent compares the result with the sources and rules and sends it back if it finds problems.
- 04
Works in parallel
Independent tasks can run at the same time, and the coordinator assembles the result.
- 05
Escalates to the supervisor
Conflicts between agents, uncertain results and actions with impact go to the person responsible.
- 06
Keeps traceability
The log shows which agent took each step, with what data and what it passed on.
How it works
How the agents work together
The steps marked in yellow are approved by a person.
- Input
- AI
- Action
- Human in the loop
- Result
- 01
Input
Complex request
For example, preparing a technical proposal for a private tender.
- 02
AI
The coordinator divides
The coordinator agent breaks down the request and allocates the tasks by role.
- 03
Action
Specialised agents work
Each uses only the tools of its role: research, calculation, drafting.
- 04
AI
The checking agent reviews
Compares the result with the sources and rules and requests corrections.
- 05
Human in the loop
The supervisor approves
The person responsible sees the result and the log and decides what goes ahead.
- 06
Result
Traceable result
The final document, with a history of each agent's contribution.
What needs to be defined for a multi-agent system
- A written definition of each role: objective, inputs, outputs and when it stops
- A permission matrix per agent: which tools and which data each one can use
- The communication protocol between agents: what information they pass on and in what format
- Checking rules and acceptance criteria for results
- Logging and monitoring infrastructure for the whole system
Possible integrations
- The company's CRM and ERP
- SharePoint, Google Drive or Confluence
- Internal databases (PostgreSQL, SQL Server)
- Web search or authorised public sources
- Microsoft Teams or Slack for approvals
- Ticketing systems (Jira, Freshdesk)
- Observability platforms for agent logs
- The company's CRM and ERP
- SharePoint, Google Drive or Confluence
- Internal databases (PostgreSQL, SQL Server)
- Web search or authorised public sources
- Microsoft Teams or Slack for approvals
- Ticketing systems (Jira, Freshdesk)
- Observability platforms for agent logs
Each integration's compatibility is checked during the analysis phase, for the systems and versions you use.
Control and transparency
What a person validates
- The final result of the process, before it is sent or used
- The roles, permissions and tools of each agent, at setup and at every change
- Situations where the agents reach contradictory conclusions
- Any external action or action with financial impact, regardless of which agent proposes it
Limits and how we manage them
- More agents mean higher running costs, latency and complexity than a single agent
- Errors can propagate between agents; checking and human supervision are mandatory, not optional
- For most processes, a single well-configured agent is enough, and we recommend it first
- A multi-agent system needs continuous monitoring and maintenance; it is not a one-off installation
- The agents do not have unlimited autonomy: they work only within the defined roles and limits
How we measure a multi-agent system
We agree the indicators together before launch and track them afterwards. The values depend on the process, the data and the volume.
- Quality of the final result
- The supervisor assesses, on a set of test cases, whether the result can be used, with or without corrections
- Each agent's contribution
- We test each role separately on known inputs, to identify where errors occur
- Effectiveness of checking
- We deliberately introduce errors into the test cases and track how many are caught by the checking agent
- Cost and time per case
- We measure model calls, duration and cost from the logs, compared with the single-agent option
- Respect for permissions
- We audit the logs to confirm that no agent used tools outside its role
How we build a multi-agent system
- 01
The “is one agent enough?” test
We analyse the process and first check whether a single agent can cover it. We only move to several if there is a clear reason.
- 02
Defining roles and permissions
We write down the roles, inputs, outputs and access matrix for each agent and for the supervisor.
- 03
Incremental build
We add the agents one at a time and test each role separately before connecting them.
- 04
Supervised pilot
We run the system on real cases with human approval at the end and adjust the roles based on the logs.
Use cases
The examples below are conceptual scenarios, not delivered implementations.
- Construction and building services
Preparing a technical proposal
One agent extracts the requirements from the tender specifications, another looks for suitable products in the internal catalogue, a third drafts the proposal, and a checker compares everything with the requirements before the manager approves it.
Conceptual example - Financial services
An analysis file for a new client
One agent gathers public information about the company, another reads the internal documents received, and a checker flags inconsistencies for the analyst who makes the decision.
Conceptual example - B2B marketing
Technical material from internal documentation
One agent extracts information from the product documentation, another drafts the material, and a checker confirms that every statement has a source before the editor's review.
Conceptual example
Frequently asked questions
What is a multi-agent system, in short?
It is an AI process split into roles: a coordinator agent receives the request and divides it, specialised agents each work with their own tools and data, a checking agent reviews the result, and a human supervisor approves it.
How do I know whether I need several agents or just one?
A single agent is usually enough when the process has a clear objective, few tools and the same permissions throughout. Several agents make sense when the stages need different access to data, when you need independent checking, or when a single agent becomes inconsistent.
Can the agents make decisions without any person involved?
Not in the systems we build. There is always a human supervisor who approves the final result and any actions with impact. The agents work only within the defined roles and permissions.
What happens when the agents disagree or are not sure?
The coordinator does not force a conclusion. The case is flagged and sent to the supervisor with both options and their sources, so that a person makes the decision.
How is sensitive data protected in a system with several agents?
Each agent gets access only to the data of its role. The agent that searches the web, for example, does not see internal documents. All exchanges between agents are logged.
Is it more expensive than a single agent?
Usually yes, because it runs more model calls and needs more monitoring. That is why we only propose it when the benefits in control and quality justify the difference.
Can’t find the answer? Write to us and we’ll reply directly.
Does your process really need several agents?
Describe the process and where it gets stuck today. We will tell you honestly whether a single agent is enough or whether a role-based architecture is worth it.
Related solutions
- AI agent for workflowsRuns multi-step processes that need decisions along the way, where fixed automation stops at the first exception, with human approval at the sensitive points.
- AI agent for researchGathers information about the market, competitors or suppliers from approved sources, with citations for every claim, while the analyst validates the conclusions.
- AI evaluation and monitoringTest sets, checks at every change, human spot-check reviews and cost tracking, so you know how your AI performs over time.