The Agentic AI design approach
Pega GenAI Agent in Pega Platform™ is the foundational building block of Pega Predictable AI™, enabling organizations to create intelligent, autonomous, and fully governed AI assistants that operate directly within enterprise workflows.
Unlike traditional generative AI tools that focus on content creation or chat, Pega GenAI Agents combine large language model reasoning with Pega deterministic rules engine, Case management, and workflow orchestration to deliver predictable, auditable, and accurate outcomes in regulated and mission critical processes.
Configured through Agent Rules and empowered by Tool Rules, Pega GenAI Agents reason step by step, plan multi-step tasks, invoke tools, retain context across interactions, collaborate with other agents by using Agent2Agent (A2A) protocol, access external data by using Model Context Protocol (MCP), and take autonomous actions. For example, advancing Cases, updating records, or triggering integrations, all while staying aligned with business logic, compliance requirements, and enterprise governance.
This approach makes Pega GenAI Agents production-ready for agentic automation in financial services, healthcare, insurance, and customer operations, bridging the gap between advanced AI capabilities and enterprise reliability.
Core principles of agentic AI
The Pega GenAI Agent Rule is a specialized interactive AI that is designed to autonomously perform tasks and provide conversational support. An effective AI agent is built on four key principles, all implemented within Pega Platform.
| Principle | Description | Pega implementation example |
|---|---|---|
| Goal orientation | The agent receives a high level objective rather than a sequence of explicit instructions. | A Case starts with the objective "Onboard new Enterprise customer ABC Corp". |
| Reasoning and planning | The agent analyzes the objective and breaks it into smaller executable steps. | The agent identifies Tasks such as creating an account, provisioning services, scheduling a welcome call, and notifying an account manager. |
| Tool use | The agent determines which tools to use to complete Tasks. | The agent uses a Connect-REST rule to call a billing API, a Data Transform to format customer data, and the Pega Email Bot to send a notification. |
| Self-correction and adaptation | The agent evaluates outcomes and adjusts actions when errors occur. | If a billing API call fails, the agent retries three times. After repeated failures, it creates a child Case for a human specialist. |
Key agentic design patterns
The technical architecture of Pega GenAI agents relies on Tool rules, which define the features that an agent uses to complete work.
Agents can perform the following actions in a Case:
- Call another agent
- Run an automated action
- Use an integration by using a buddy
- Start a Case
- Retrieve information from a Data Page
- Run a Flow Action
The architecture of Pega GenAI Agents is designed to be extensible and supports advanced protocols that enable broader integration and enhanced functionality. These protocols allow agents to participate in distributed, goal-oriented workflows that extend beyond Pega Platform.
The Agent2Agent (A2A) protocol enables communication between Pega GenAI Agents and external agents to coordinate and complete tasks collaboratively. This ability supports the design of complex workflows that span multiple AI systems.
The Model Context Protocol (MCP) further extends an agent’s abilities by providing access to external data sources and contextual information. This additional context complements native Pega GenAI features and enables agents to make more informed decisions during run time.
As a Lead System Architect (LSA), your role goes beyond using agents: you are ALSO responsible for designing and orchestrating them. Within Pega Platform, you can apply established design patterns to define how agents interact, integrate with external systems, and contribute to scalable, secure solutions.
AI as a digital case worker
This pattern positions a Pega GenAI agent as the primary Case owner. The agent completes the initial tasks, gathers data, and creates subcases only when human expertise or approval is required.
- Best for: Repetitive, complex processes such as claims processing, customer onboarding, or service provisioning.
- LSA responsibility: You define the Case Type structure and the conditions that trigger a human review, such as a claim value threshold or a VIP customer attribute.
Hierarchical agent pattern (manager or worker agents)
This pattern uses a 'manager' agent to break a complex goal into smaller sub goals and delegate each to specialized 'worker' agents.
- Best for: Multi-layered, large-scale projects, such as marketing campaigns. For example, A Marketing Campaign Manager agent can delegate responsibilities to multiple worker agents, such as:
- A Social Media Content agent to generate and schedule posts.
- An Email Campaign agent to design and execute targeted outreach.
- A Performance Analytics agent to evaluate engagement and outcomes.
- LSA responsibility: You define interaction protocols, data contracts, tool availability, and result aggregation logic.
Human in the loop supervisory pattern
This pattern ensures oversight for high risk or sensitive processes. The agent provides recommendations, and a human approves the action.
- Best for: High-risk tasks such as financial decisions or software release approvals.
- LSA responsibility: You design approval interfaces, define SLAs, specify escalation paths, and ensure that all decisions are auditable.
For a Lead System Architect, Pega GenAI Agents represent more than a new Pega Platform feature; they require deliberate architectural leadership. Establishing clear, reusable agent design patterns is essential to delivering autonomous enterprise applications that operate reliably and scale with confidence.
In the emerging era of agentic AI, disciplined architecture is what transforms intelligent agents from isolated capabilities into enterprise grade systems.
Check your knowledge with the following interaction:
This Topic is available in the following Module:
Want to help us improve this content?