Lesson 1 of 6
Two protocols, two jobs: MCP and A2A
7 min
Listen to this lesson
Two open protocols now decide how AI agents reach enterprise systems. The Model Context Protocol (MCP) connects an AI application to tools and data. The Agent2Agent protocol (A2A) connects one agent to another so they can hand off work. SAP's own API Policy overview puts it simply: MCP connects agents to systems and data, A2A lets agents collaborate across systems. Get this distinction right and most architecture questions in this course become easier.
MCP: an agent uses tools
MCP standardizes how an AI application discovers and calls capabilities. The application that hosts the model is the host; inside it, a client keeps a connection to each server; each server offers capabilities. Messages use JSON-RPC 2.0. A server can offer three kinds of capability, which SAP Integration Suite also calls MCP server primitives:
- Tools: functions the model can call to do something, each with an input schema (for example, "get sales order status" with an order number).
- Resources: data and context the model or user can read, identified by URIs (for example, a product catalog document).
- Prompts: reusable, parameterized message templates that guide how a task is done.
Clients can also offer features to servers. In the current specification (revision 2026-07-28) that is elicitation, a way for a server to ask the user for missing information. That revision also made the base protocol stateless, with self-contained requests. Earlier revisions, such as 2025-11-25, used a session with an initialize handshake. This matters in practice: SAP Integration Suite documents support for both revisions, and you connect a client using the revision it was built on. Remote MCP servers use the Streamable HTTP transport.
A2A: an agent delegates a task
A2A standardizes how one agent asks another to do a piece of work and gets the result back, without either seeing the other's internals. Each A2A server publishes an Agent Card, a JSON document (conventionally at /.well-known/agent-card.json) that describes the agent's skills, endpoint and authentication requirements. Work is a Task with a lifecycle: states such as submitted, working, input-required, completed, failed, canceled and rejected. Agents exchange Messages made of Parts (text, files, structured data), and results come back as Artifacts. Long tasks can stream updates or notify a webhook.
| Question | MCP | A2A |
|---|---|---|
| What is on the other side? | A tool or data source with a schema | Another agent with its own reasoning |
| Unit of work | A tool call or resource read | A task with states, messages and artifacts |
| Who plans the steps? | The calling agent plans every call | The remote agent plans its own steps |
| Discovery | The client lists the server's tools, resources and prompts | The client reads the Agent Card |
| Typical SAP example | An agent reads sales order 1000457 through a governed tool | Joule delegates "assess delivery risk" to a specialist agent and gets a report back |
| Version to know (Oct 2026) | Spec revision 2026-07-28; 2025-11-25 still widely used | A2A 1.0 released; SAP's documented Joule integration uses 0.3.0 |
MCP and A2A answer different questions; most real designs use both.
Who governs the standards
A2A was created by Google and donated to the Linux Foundation; its technical steering committee includes SAP alongside AWS, Cisco, Google, IBM Research, Microsoft, Salesforce and ServiceNow, and the project site announces that A2A is joining the Agentic AI Foundation. MCP sits in the Agentic AI Foundation as well, and SAP's overview deck lists SAP as a gold member co-chairing the identity and trust working group. The point for architects: both protocols are open, multi-vendor and still moving, so pin the revision you build against.
- Does the agent need to read data or perform one well-defined operation?Yes → an MCP tool or resource fits · No → continue.
- Should another agent own the plan, the steps and the business rules?Yes → delegate with A2A · No → keep the plan in your agent and call tools.
- Is the other side a system API rather than an agent?Yes → MCP (or plain APIs for deterministic integration) · No → A2A.
- MCP connects an AI application to tools, resources and prompts on MCP servers over JSON-RPC 2.0.
- A2A lets a client agent hand a task to a remote agent, discovered through its Agent Card, and get artifacts back.
- With MCP the calling agent plans each step; with A2A the remote agent owns its own plan.
- Both are open and evolving: MCP revision 2026-07-28 is stateless, and A2A has reached 1.0 while SAP documents 0.3.0.
Check your understanding · not graded
Lesson check
1. Which MCP capability lets a model perform an action such as reading an order's status?
2. Where does an A2A client conventionally find a remote agent's capabilities?
3. Joule should hand "assess supplier delivery risk" to a specialist agent and receive a report. Which protocol fits?
4. What changed in the MCP base protocol with revision 2026-07-28?