Chapter 9: Multi-Agent Patterns
As tasks grow more complex, single agents hit limits. Multi-agent patterns coordinate multiple specialised agents working together.
Pattern 6: Delegation
A primary agent receives a request and delegates sub-tasks to specialist agents, then synthesises their outputs.
When to use: When tasks span multiple domains of expertise.
In the Pragmatix Advisory Portal, a complex question about implementing an agentic system might involve the Agentic AI Advisor (architecture), the Digital Governance Advisor (policy), and the Cyber Security Advisor (risk). A coordinating agent could delegate to each and synthesise.
Pattern 7: Specialisation
Rather than one generalist agent, you build multiple specialists — each with focused DNA for a specific domain.
Benefits:
- Deeper expertise in each domain
- Simpler, more maintainable agent configurations
- Clearer boundaries and responsibilities
When to use: When expertise requirements vary significantly across domains.
The Pragmatix Advisory Portal uses this pattern — separate advisors for Digital Governance, Solution Architecture, Cyber Security, and Agentic AI rather than one "everything" advisor.
Pattern 8: Pipeline
Agents are arranged in sequence, each performing a specific transformation. The output of one becomes the input to the next.
When to use: When a task naturally decomposes into sequential stages with clear handoffs.
Pattern 9: Consensus
Multiple agents independently work on the same problem, then their outputs are compared and reconciled.
When to use: High-stakes decisions where you want multiple perspectives or redundancy.
Pattern 10: Supervisor
A supervisory agent oversees the work of other agents — assigning tasks, monitoring progress, handling exceptions, and ensuring quality.
When to use: Complex workflows where coordination and oversight are critical.