Agent-to-Agent Protocol (A2A): What It Is in 2026

The Agent-to-Agent Protocol (A2A) is an open interoperability standard for AI agents launched by Google on April 9, 2025 and contributed to the Linux Foundation in June 2025. By the end of 2025, A2A had 150+ partner organizations including all major hyperscalers, with version 0.3 (gRPC support, security card signing) released July 31, 2025 and a stable interface released for enterprise adoption.

The protocol solves a different problem from ACP (which is checkout-focused) or UCP (which is commerce-journey). A2A sits at a layer below: how do two AI agents from different vendors talk to each other? An agent built on Anthropic Claude and an agent built on Google Gemini both need to coordinate when they're working on the same multi-step task. A2A standardizes that conversation.

For commerce specifically, A2A is the substrate that lets an enterprise procurement agent talk to a vendor's sales agent, or a customer's personal shopping agent talk to a retailer's catalog agent, without bespoke integrations. UCP, AP2, and Visa TAP all explicitly cite A2A compatibility as part of their interop story.

A2A standardizes three things across agent vendors:

Discovery. An agent advertises its capabilities through a standardized agent card (a JSON descriptor of what it can do, what it expects as input, what it returns). Other agents read agent cards to find collaborators without prior integration.

Message format. Agents exchange typed, versioned messages. v0.3 added gRPC support alongside HTTP for higher-throughput agent-to-agent flows. Tasks have explicit lifecycle states (pending, in-progress, completed, failed) so calling agents can track progress.

Security card signing. Agent cards are cryptographically signed. A receiving agent can verify the calling agent is who it claims to be before delegating sensitive operations.

The Linux Foundation handover (June 2025) was the inflection point. Before that, A2A was Google's. After, it became the kind of vendor-neutral standard that hyperscalers (AWS, Azure, GCP), enterprise platforms (Salesforce, ServiceNow, SAP), and AI agent vendors (Anthropic, OpenAI) could co-govern.

The four open protocols sit at different layers:

Practical example: a customer's ChatGPT-based shopping agent (running on ACP for the merchant flow) needs to coordinate with a retailer's loyalty-program agent (running on a different vendor stack). A2A is what makes that handoff work without a one-off integration. The retailer doesn't need to integrate with every customer-side agent vendor; A2A is the common ground.

Most retailers will not implement A2A directly. Their AI agents (customer-service agents, internal-team agents, vendor-procurement agents) get A2A support through the platform they're built on, the same way most retailers get HTTP support through their web framework.

What matters at the retailer level:

1. Lower lock-in. A retailer building agent infrastructure today doesn't have to bet on one AI vendor. A2A lets them swap underlying agent vendors as the market shifts without rebuilding integrations.
2. Cross-surface consistency. A retailer's internal agent (for store associates, e.g. Walmart's 2.1 million associate AI training) and external agents (for customer-facing AI shopping) can interoperate via A2A without the retailer building bridges.
3. Foundation for AP2 and UCP. The payment-layer (AP2) and commerce-journey-layer (UCP) protocols both rely on A2A interop primitives. Retailers adopting either inherit A2A compatibility.

A2A is the kind of infrastructure standard retailers can mostly ignore at the integration level but should track because it shapes which agent platforms are bettable. Linux Foundation governance + 150+ org backing means it's not going away.