Divide work by role
One agent researches, a second reviews and a third prepares the handover. Each role receives only the capabilities and information it genuinely needs.
Coordination has a cost and needs to deliver value
Assignments, shared case context, conflict, stop conditions and human escalation need deliberate design. A multi-agent system makes sense when separation improves quality, control or extensibility.
- clear ownership
- structured handovers
- separate review roles
- human conflict resolution
Every role needs a contract, context and owner
Before technical implementation, each digital role receives a defined input, permitted output and accountable business owner. Handovers need a shared, bounded data format instead of uncontrolled conversational history.
- unambiguous assignment
- minimum necessary tools
- bounded case context
- reviewable handover outcome
- human process owner
Divided work must not multiply risk
More roles create additional identities, permissions, messages and failure paths. Access is therefore granted per role, while especially consequential actions remain separated or gated by approval.
A reviewing agent is not an independent expert review where both roles share the same blind spots, sources or models.
The whole system is tested at handovers and under disruption
Tests cover not only each role but also lost context, conflicting results, repeat loops, unavailable tools and unauthorised expansion of an assignment. Stop and escalation criteria need to function in those conditions too.
A change to one role may affect downstream roles and therefore requires renewed end-to-end evaluation.
When several agents are justified
Separation is useful where it makes accountability boundaries, different access rights or independent work stages clearer. For a short linear flow, it may only add complexity and operating effort.
The smallest architecture that fulfils the task and its controls reliably is the better starting point.
Source note
The architecture principles are derived from NIST lifecycle risk management and BSI guidance on generative AI. Primary sources reviewed on 11 August 2026.
Primary sources
Official sources, editorially checked on 11 August 2026.
- AI Risk Management FrameworkNational Institute of Standards and Technology (NIST)
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile (NIST AI 600-1)National Institute of Standards and Technology (NIST)
- Generative AI Models: Opportunities and Risks for Industry and AuthoritiesBundesamt für Sicherheit in der Informationstechnik (BSI)