Chapter 13: Patterns in Practice
Patterns are building blocks, not blueprints. Real-world agentic systems typically combine multiple patterns.
Example: An Autonomous Accounts Payable System
Consider a system that processes invoices automatically:
| Pattern | Application |
|---|---|
| Specialisation | Separate agents for invoice processing, vendor matching, and approval routing |
| Pipeline | Invoices flow through extraction → validation → matching → approval |
| Approval gates | Invoices above threshold require human approval |
| Confidence-based escalation | Uncertain matches flagged for review |
| RAG | Agents retrieve vendor contracts and payment terms |
| Supervisor | Coordinating agent monitors pipeline, handles exceptions |
For another real-world example of patterns in combination, see the PIMS asset register case study in Chapter 4 (Runtime Architectures). It demonstrates Tool Use (Pattern 2) combined with Approval Gates (Pattern 11), with a clear graduation path to Confidence-Based Escalation (Pattern 12): a practical illustration of how organisations often build agentic capabilities within existing applications before recognising them as such.
Designing Combined Pattern Systems
Designing such a system requires:
- Understanding the available patterns
- Assessing your risk tolerance and quality requirements
- Making deliberate choices about where humans stay in the loop
This is where the Pragmatix Agentic AI Maturity Assessment provides value, helping you understand your current state, identify appropriate patterns for your context, and chart a pragmatic path forward.
Looking Ahead
Part 3 covers what it takes to move these patterns from prototype to production: evaluation, security, cost management, and operations.