A2A and MCP in Pega Agentic AI
As you begin designing solutions with agentic features in Pega Infinity™, architectures evolve quickly from a single autonomous agent to sophisticated multi-agent autonomous systems composed of specialized Agents working together.
When multiple agents work together, they must exchange context, delegate tasks, and coordinate outcomes reliably while maintaining autonomy and meeting enterprise governance requirements.
Pega Platform™ aligns with two open, industry-standard protocols for this purpose:
- Agent2Agent (A2A) protocol for peer-to-peer agent collaboration.
- Model Context Protocol (MCP) for secure, governed access to tools and data.
Understanding when and how to apply these protocols is essential for designing scalable, interoperable, enterprise-ready multi-agent autonomous systems.
Agent2Agent protocol
The A2A protocol is an open, industry-standard communication framework that enables secure collaboration between autonomous AI agents, regardless of vendor, framework, or platform.
Introduced by Google and now governed as an open-source initiative, A2A addresses a key limitation of agentic AI: no single agent can possess all the specialized expertise or data required for complex enterprise tasks. By providing a standardized "common language," A2A helps agents discover one another through secure agent cards that describe features, endpoints, and authentication requirements. Agents can then delegate subtasks, exchange information, and coordinate actions without exposing internal logic or requiring custom integrations.
The A2A protocol supports collaborative, goal-oriented processing across multiple agents. In Pega Agentic AI, its technical features include:
- Bidirectional integration: A Pega-defined Agent can invoke an external agent, and an external agent can invoke a Pega Agent.
- Orchestration architecture: The Pega Agentic Process Fabric™ includes an intelligent registry that maintains visibility into available Agents and workflows.
- Fabric control Agent: An AI-driven control Agent selects the most suitable Agent from the registry and uses the A2A protocol to delegate tasks and receive completion results.
In practice, A2A is used to orchestrate multi-agent workflows in enterprise scenarios that require predictability, governance, and interoperability. For example, during an expense report review Case, a Pega Agent can invoke an external fraud-detection Agent through A2A. The external Agent validates transactions against vendor blocklists, identifies currency exchange anomalies, and returns its findings so that the Agent can continue the Case with enhanced decisioning intelligence.
By adopting the A2A protocol, enterprises can move beyond monolithic AI implementations toward flexible, ecosystem-wide networks of autonomous agents that improve accuracy and efficiency while reducing integration overhead and vendor lock-in.
Model Context Protocol
The MCP is an open, industry-standard protocol introduced by Anthropic and governed by the Agentic AI Foundation. MCP acts as a universal adapter for AI agents by providing a secure, standardized client-server architecture. With this architecture, AI applications, such as large language models and autonomous Agents, can connect dynamically to external data sources, tools, APIs, databases, content repositories, and workflows.
MCP defines a consistent interaction model for communication between AI Agents and enterprise systems by using a host-client-server pattern:
- The AI agent (host) initiates a request to perform an action.
- The client translates the request into the MCP standard.
- The server exposes approved tools or data through a secure interface.
Unlike fragmented custom integrations that require bespoke code for each system, MCP replaces them with a single, interoperable protocol. An MCP client embedded in the AI system communicates with MCP servers hosted alongside enterprise data or tools to retrieve real-time context, run actions, or access structured information without duplicating data or exposing sensitive internal logic. By grounding agent reasoning in authoritative, live sources, this approach helps reduce hallucinations while supporting scalable, two-way interactions across enterprise environments.
In practice, MCP is used to extend agent features in scenarios where access to up-to-date or specialized external context is critical for reliable decision-making and automation. MCP allows Pega Agents to safely access enterprise systems, run tools, retrieve data, and perform governed operations without embedding custom logic or integrations in the Agent. This approach supports secure, reusable, and centrally governed access to enterprise features.
Architectural decision guide: A2A versus MCP
The A2A and MCP are both essential to agentic AI architectures, but they serve fundamentally different architectural purposes. Selecting the appropriate protocol requires considering not only the immediate task, but also the full life cycle and nature of the business process being automated. The following tables provides guidance for choosing the right protocol for a given design scenario:
| Feature | A2A | MCP |
|---|---|---|
| Purpose | Designed for multi-agent collaboration. Enables autonomous agents to discover features, share context, negotiate responsibilities, and collaborate to achieve goals. | Designed for tool and data access. Provides a standardized way for an AI model to securely invoke tools, retrieve data, or perform actions within enterprise systems. |
| Primary use case | When multiple agents, either Pega-defined or external, must work together for cross-platform workflows, orchestration, and intelligent delegation. | When a single agent requires controlled access to information, services, or tools, such as querying systems, retrieving data, or executing well-defined operations. |
| Interaction model | Agent-to-agent communication with peer-level collaboration and capability exchange. | Model for tool-to-server interaction through a standardized interface. |
| Complexity and scope | Dynamic and flexible. Supports planning, negotiation, capability discovery, and multi-step goal processing. | Structured and controlled. Focuses on exposing secure, permissioned functions and data to an AI model. |
| Conceptual analogy | Comparable to a team of specialists collaborating, where each agent contributes distinct expertise. | Comparable to a universal adapter that allows an AI model to connect to existing tools and data sources. |
| Pega usage | Enables Pega Agents to collaborate with other Pega Agents or external AI agents, supporting cross-system workflows, IT service management automation, fraud analysis, and agentic customer service. | Pega Agents can safely access enterprise data and tools without custom integrations, supporting record retrieval, service invocation, and governed execution of actions. |
| When to choose | The process spans multiple agents or systems, requires intelligence. | A single agent requires access to tools or data under strict conditions. |
The A2A and MCP patterns represent fundamental architectural shifts in how AI is integrated into enterprise processes. They represent complementary architectural approaches, not competing standards. A2A provides the collaboration framework required for multi-agent autonomous systems, so agents can work together as a cohesive team. MCP acts as a universal access layer, and agents can securely and reliably interact with enterprise tools and data.
For Lead System Architects, the ability to apply these protocols strategically is critical to building solutions that are innovative, scalable, and governable. Knowing when to orchestrate a team of autonomous agents versus when to equip a single agent with controlled system access is a cornerstone of effective agentic AI design on Pega Platform.
Check your knowledge with the following interaction: