One of those hops is MCP. Which one, and what breaks if you pick the wrong one?

Your case agent needs a customer's exposure from the core banking API. It also needs a call from another team's fraud agent.

The 52-second version.

1. MCP is the tool boundary.

MCP is an open protocol, introduced by Anthropic and governed by the Agentic AI Foundation. It gives an agent one standard way to reach tools and data. The core banking call is the MCP hop.

2. A2A is the peer boundary.

A2A is an open standard for secure collaboration between autonomous agents. It holds regardless of vendor, framework or platform. The fraud agent the other team owns sits on the far side of A2A.

3. MCP has three roles. Name all three.

The agent is the host and raises the request. The client translates it into the MCP standard. The server exposes only approved tools and data. Candidates who skip the client have missed where control actually sits.

4. A2A is only as good as the registry behind it.

The Pega Agentic Process Fabric keeps an intelligent registry of available agents and workflows. Strip that visibility out and agent discovery becomes point-to-point integration wearing a new name.

5. The control Agent selects. It does not orchestrate.

An AI-driven control Agent picks the most suitable agent from the registry. It delegates over A2A, then takes back the completion result. Selects and delegates, not orchestrates. That distinction is the reason A2A exists.

6. Get it backwards and you lose the thing you needed.

Pega's own rule is short. A2A when the process spans multiple agents or systems. MCP when one agent needs tools or data under strict conditions. Model that fraud agent as an MCP tool and you still get your answer. You lose the negotiation, the registry entry and the audit trail that came with it.

Lesson

MCP answers what your agent may touch. A2A answers who else it may ask. Two questions, two protocols, and most real designs need both.

Book a Discovery Call