Search

What are you looking for?

Search our services, use cases and practical insights.

Enter at least 2 characters

Popular starting points

Agents and automation

When do multiple AI agents make sense?

Multiple agents can divide complex work, but they add hand-offs, latency, cost and failure modes. They are an architecture tool for clearly separable tasks, not an automatic quality multiplier.

The short answer

A multi-agent system coordinates several specialised agents that perform subtasks or hand results to one another. It is useful when roles, contexts or tools genuinely need to be separated and the hand-offs can be controlled objectively.

In brief

  • Multiple agents are worthwhile only when work divides naturally into accountable, testable subtasks.
  • An orchestrator or defined process decides which agent works when and how results are combined.
  • Every extra hand-off can lose context, amplify errors and increase latency and cost.
  • A single agent or workflow should be evaluated on the same cases as a simpler baseline.

How can multiple agents work together?

In a manager pattern, one central agent decomposes the assignment, calls specialists and combines their results. In a hand-off pattern, one agent transfers the case to another. Parallel patterns ask several agents to handle independent perspectives that are then combined by rules or an accountable decision maker.

Roles should reflect technical boundaries such as different data domains, tools or evaluation criteria. Several named personas with the same context and permissions mostly add complexity.

How can multiple agents work together?
PatternControlUseful whenMain risk
Manager and specialistsOne central agent assigns and combines subtasksDecomposition varies while specialist outputs remain testableThe manager loses evidence or combines contradictions incorrectly
Hand-offOne agent transfers the complete case to the next roleOwnership and access boundaries change clearlyIntent, context or open issues disappear during transfer
ParallelSeveral roles work independently before results are combinedPerspectives are independent and time saving is measurableDuplicate work and conflicting results without a clear decision
Workflow with agent stepsSoftware fixes sequence and hand-off contractsThe overall process is stable but individual steps need flexible AIOverly rigid interfaces or unnecessary model calls

When does division provide a real advantage?

Division can help when one context becomes too large, teams have separate access boundaries or subtasks should be scaled and evaluated independently. Each agent then receives only the information and tools needed for its role.

  • Subtasks have distinct objectives and testable outputs
  • Data or permission boundaries should remain technically separated
  • Specialist contexts would overload a single agent
  • Parts can run in parallel and demonstrably save time

Which additional risks appear?

A wrong intermediate result can be treated as fact and amplified by later agents. Hand-offs can lose sources, uncertainty or the original user intent. The number of possible paths also grows, making evaluation and troubleshooting harder.

  • Duplicate or conflicting actions caused by unclear ownership
  • Loss of provenance and uncertainty at interfaces
  • Unpredictable loops, latency and model spend
  • A larger attack surface through additional tools and messages
  • Unclear accountability for the combined final result

How should the architecture be chosen and evaluated?

Start with a simple baseline: a workflow or one agent. Add a role only when a concrete boundary or measurable quality gain justifies it. Every hand-off should use a defined schema containing output, evidence, uncertainty and status.

  1. Step 1

    Justify the split

    For each role, record the problem it solves compared with a simpler design.

  2. Step 2

    Define contracts

    Specify inputs, outputs, evidence, failures and stop conditions for every hand-off.

  3. Step 3

    Evaluate separately

    Test individual agents and their permissions in isolation on representative cases.

  4. Step 4

    Compare the complete system

    Measure quality, latency, cost and troubleshooting effort against the baseline.

Example from day-to-day business

Example: preparing a tender response

A coordinator separates a requirement list into defined domains. A product agent reads approved product data, a compliance agent checks authoritative evidence and a calculation service returns structured prices. Every part includes sources and open issues. An accountable person reviews the combined dossier and approves it. No agent independently invents or sends customer-specific commitments.

What to remember

Use multiple agents only when subtasks, permissions and hand-offs can be separated clearly and the more complex architecture performs measurably better than a simpler alternative.

Sources and further reading

These primary sources provide further detail on definitions, technical foundations or responsible use.

Content reviewed

Would you like to apply this to your situation?

Together, we clarify what makes sense for your process, data and systems – in plain language and without unnecessary complexity.

Discuss Your Project