Concept card

Agent-to-Agent (A2A) Communication via MCP

As of 2026-07-24T14:00:00Z

Edited by Cédric-Olivier Mâry, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

What is Agent-to-Agent (A2A) Communication via MCP?

MCP's real value for SAP is eliminating the bespoke REST contract each agent-to-agent integration used to require — a Joule orchestrator now discovers a third-party agent's tools dynamically via a standard tools/list response.

What it is

Agent-to-agent (A2A) communication via the Model Context Protocol (MCP) is the pattern by which two AI agents — possibly built on different large-language-model runtimes, owned by different teams, or running in different cloud tenants — exchange structured requests and responses over a standardized transport, using JSON-RPC 2.0 as the wire format. MCP was open-sourced by Anthropic in late 2024 and has since become the closest thing the industry has to a common interoperability layer for agent ecosystems, including SAP's BTP-hosted Joule agents calling out to specialist third-party agents, and third-party agents calling back into Joule.

Why it matters

Before a shared protocol existed, every agent-to-agent integration was a bespoke contract. A Joule procurement agent that needed a supplier-risk score from a model built on a different platform required both engineering teams to agree on a proprietary REST schema, maintain its versioning by hand, and re-implement authentication on both sides — for every new pairing. MCP standardizes the handshake instead: the calling agent, the MCP client, sends an initialize request, discovers what the other agent can do through a tools-list response, and then invokes a specific capability with a tools-call request. Authentication rides on standard OAuth 2.1 bearer tokens rather than a custom scheme. Neither team writes bespoke glue code, and a new specialist agent can be plugged into an existing orchestrator without a redesign.

How it works

Why it matters in practice

Before MCP, integrating a Joule agent with a third-party agent (e.g. a Databricks risk-scoring agent) meant both teams building and versioning a proprietary REST contract
Capability advertisement via tools/list is machine-readable, letting LLM planners discover available tools at runtime instead of relying on hardcoded integrations
Production inter-service calls across BTP subaccounts use Streamable HTTP/SSE; stdio is for same-process or local development only

Key points

MCP (Model Context Protocol, Anthropic 2024) standardises agent-to-agent calls via JSON-RPC 2.0 — tools/list for capability discovery, tools/call for invocation.
Two transports: stdio (same-process, local dev) and Streamable HTTP/SSE (production inter-service, BTP subaccount-to-subaccount).
MCP enables dynamic tool discovery: the calling LLM planner receives tools/list at runtime, removing hardcoded routing logic from orchestrator prompts.
MCP adds ~5–15 ms per call vs lean REST; irrelevant for conversational latency, relevant at > 10,000 calls/minute throughput.
Security: OAuth 2.1 bearer tokens mandatory in production; scope claim must be validated per-tool on the server side.
Prompt injection risk: malicious tools/list descriptions can hijack the calling LLM planner — apply a trust registry and strip HTML/markdown from tool descriptions.
SAP BTP AI Core tool_execution framework maps to MCP tool pattern; no native SAP MCP SDK GA as of 2026 but community adapters exist.
Agent-to-Agent (A2A) Communication via MCP is mastered only when it changes a named buyer decision.
Start with the semantic contract and control model before demonstrating the tool.
Use current SAP, analyst, study, KG, and news signals as evidence, not decoration.

Terms used on this page

MCP (Model Context Protocol): Open JSON-RPC 2.0 protocol for agent-to-agent communication, open-sourced by Anthropic November 2024. Defines tools, resources, and prompts as first-class primitives.
tools/list: MCP method by which a server advertises its available tools to any MCP client. The response feeds directly into a calling LLM's function-calling context.
tools/call: MCP method by which a client invokes a specific tool on the server, passing structured arguments and receiving a structured result.
Streamable HTTP/SSE: MCP's production transport: HTTP POST for requests, Server-Sent Events for streaming responses. Used for cross-process and cross-network agent calls.
Trust Registry: A pre-approved allow-list of MCP server URIs that an MCP client will invoke. Prevents prompt injection via malicious servers.
/.well-known/mcp.json: Community-adopted discovery file that advertises an organisation's MCP server URI, available tools, and capability catalogue to AI crawlers and agent directories.
Decision owner: The accountable person who accepts the trade-off and funds the next action.
Semantic contract: The shared definition of business terms, metrics, entities, and access rules used by tools and teams.

Sources

Full card available to members. What the full card adds: the full decision framework · the SAP vs Snowflake / Databricks / Fabric comparison · the common pitfalls and their fix · the cheat sheet · the architecture schemas · the code blocks · the facts worth quoting.

Open in the app →