Concept card

Agent-to-Agent (A2A) Protocol — Cross-Vendor Agent Orchestration

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) Protocol — Cross-Vendor Agent Orchestration?

A2A standardizes the agent-to-agent boundary — letting one autonomous agent discover and delegate a task to another regardless of vendor or model — complementing MCP, which only standardizes agent-to-tool calls.

What A2A Is

Agent-to-Agent (A2A) is an open communication protocol, announced by Google in April 2025 with more than fifty founding partners including SAP, Salesforce, ServiceNow, Workday, Atlassian, MongoDB, and PayPal. Governance passed to the Linux Foundation under a dedicated A2A Project, with reference software development kits maintained in Python, Java, and dot-NET. The protocol answers a question that autonomous business agents raised the moment more than one vendor started building them: how does an agent built by one company discover, negotiate with, and delegate work to an agent built by a different company, running a different model, on a different platform, without either side exposing its internal application programming interfaces?

How It Works

Why it matters

Announced by Google in April 2025 with 50+ founding partners including SAP, Salesforce, ServiceNow and Workday, now governed by the Linux Foundation since June 2025.
The Agent Card at /.well-known/agent.json lets an orchestrating agent check a peer's capabilities and auth scheme before delegating a task.
Tasks carry a lifecycle (submitted → working → input-required → completed/failed/canceled) with SSE streaming via tasks/sendSubscribe for long cross-vendor interactions.

Key points

A2A protocol — announced by Google in April 2025 with 50+ founding partners including SAP; transferred to Linux Foundation A2A Project governance in June 2025; SAP was a day-one signatory, confirming A2A as a BDC/Agent Hub integration standard.
Boundary it standardises — agent-to-agent (cross-vendor); MCP (C224) standardises agent-to-tool (within a vendor's domain). The two protocols are complementary, not competing: MCP inside each agent's boundary, A2A between agents.
Four primitives — Agent Card (/.well-known/agent.json: capabilities, skills, modalities, auth scheme) · Task (lifecycle: submitted → working → input-required → completed/failed/canceled) · Message (multi-modal payload: text, file, structured data) · Streaming (server-sent events on tasks/sendSubscribe for long-running interactions).
Canonical SAP scenario — Joule Agent (BTP) detects data quality issue → delegates cleanup Task to Salesforce Agentforce via A2A → Agentforce issues A2A Task to ServiceNow flow agent for ITSM record. No shared internal APIs; all governed by SAP AI Agent Hub (C211).
SAP AI Agent Hub (C211) is the A2A governance plane — it registers authorised peer agents, logs all delegations, enforces rate limits and policies, and provides the audit trail. A2A is the wire; Agent Hub is the control tower.
Authentication negotiated per task — OAuth 2.0, API keys, mTLS. For regulated SAP customers, mTLS is the recommended pattern: mutual authentication between BTP and peer agents satisfies EU AI Act Art. 9 documentation requirements.
Task lifecycle control — tasks/send (fire-and-forget), tasks/get (status poll), tasks/cancel. The 'input-required' lifecycle state enables human-in-the-loop pauses for high-stakes delegated decisions before a peer agent proceeds.
Agent Card is the discovery contract — an orchestrating agent reads the peer's /.well-known/agent.json before issuing any Task; the Card declares whether the peer supports the required skill, modality, and auth scheme. Incomplete or missing Agent Cards prevent A2A integration entirely.
A2A vs proprietary orchestration (e.g. LangChain multi-agent, AutoGen) — A2A enables cross-vendor production deployments without trusting the peer's internal implementation; proprietary frameworks require a shared runtime, which is impractical across SAP, Salesforce, and ServiceNow estates simultaneously.
The Linux Foundation governance structure ensures vendor-neutral evolution of the A2A spec — no single vendor (including Google) can unilaterally change the protocol in ways that break SAP's Agent Hub implementation. This is the governance property that distinguishes A2A from proprietary API integrations.

Terms used on this page

Agent Card: Public JSON descriptor at /.well-known/agent.json that declares an A2A agent's capabilities, skills, supported message modalities (text, audio, structured data), and authentication scheme. The discovery contract a calling agent reads before initiating any Task delegation.
Task (A2A): The unit of delegated work one A2A agent submits to another. Tasks are submitted via tasks/send (fire-and-forget), polled via tasks/get, or cancelled via tasks/cancel. They progress through a lifecycle: submitted → working → input-required → completed/failed/canceled.
input-required lifecycle state: A Task lifecycle state indicating that execution is paused because the receiving agent requires a human decision or a peer confirmation before proceeding. The primary human-in-the-loop control point in A2A workflows.
tasks/sendSubscribe: A2A streaming method using server-sent events (SSE) that delivers real-time progress updates on a Task without requiring the calling agent to poll — used for long-running cross-vendor interactions such as multi-step data quality remediations.
A2A Project (Linux Foundation): The vendor-neutral governance body for the A2A specification since June 2025; ensures no single vendor can unilaterally modify the protocol in ways that break existing implementations. SAP, Google, Salesforce, ServiceNow are all founding members.
MCP (Model Context Protocol): Anthropic-originated protocol (C224) that standardises the agent-to-tool boundary — a single agent calling a server's tool catalogue to retrieve data or execute an action. Complementary to A2A: MCP operates inside one agent's boundary; A2A operates between agents across vendor boundaries.
SAP AI Agent Hub (C211): The governance plane above A2A in the SAP BTP/BDC stack: registers authorised peer A2A agents, logs every Task delegation, applies rate limits and policy controls, and generates the audit trail for compliance reporting. A2A is the wire; Agent Hub is the control tower.
mTLS (mutual TLS): Mutual Transport Layer Security: an authentication pattern where both the calling agent and the receiving agent present X.509 certificates to each other, establishing verified mutual identity. The recommended A2A authentication scheme for regulated SAP customers, providing a cryptographically verifiable connection record that satisfies EU AI Act Art. 9 documentation requirements.

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 →