The agent protocol stack in 2026: MCP, A2A, AP2, UCP and ACP
Which protocol does what, who governs it, what is production-ready in October 2026, and where each one lands in an enterprise integration landscape.
Two years ago every agent framework invented its own way to call tools and talk to other agents. By late 2026 that argument is largely over. The industry has settled on two horizontal protocols, both now governed by the same neutral foundation, with a small number of vertical protocols for commerce and payments built on top. For integration teams this is good news: the stack finally looks like something you can put on an architecture slide and govern.
The stack at a glance
| Protocol | Connects | Origin | Governance (Oct 2026) | Status |
|---|---|---|---|---|
| MCP (Model Context Protocol) | Agent → tools, data, prompts | Anthropic, Nov 2024 | Agentic AI Foundation (Linux Foundation) | Current spec 2026-07-28; Tier 1 SDKs in TypeScript, Python, Go and C# |
| A2A (Agent2Agent) | Agent ↔ agent | Google, Apr 2025 | Linux Foundation project since Jun 2025; joined the Agentic AI Foundation Aug 2026 | v1.0 (Mar 2026), v1.0.1 (May 2026) |
| AP2 (Agent Payments Protocol) | Payment authorization evidence | Google, 2025 | Open specification; built as an A2A extension | Early adoption with payment networks and PSPs |
| UCP (Universal Commerce Protocol) | Agent ↔ merchant: discovery, catalog, cart | Google and Shopify, Jan 2026 | Open specification with partner coalition | Early adoption in retail |
| ACP (Agentic Commerce Protocol) | Checkout inside an AI surface | OpenAI and Stripe, Sep 2025 | Vendor-maintained open specification | In production for chat-based checkout |
Naming trap: IBM's earlier Agent Communication Protocol, also abbreviated ACP, merged into A2A in August 2025 and is no longer developed. When someone says "ACP" in 2026 they almost always mean the OpenAI/Stripe commerce protocol.
MCP: the agent-to-tool layer
MCP standardizes how an AI application (the host, via an MCP client) discovers and calls capabilities exposed by an MCP server: tools the model can invoke, resources it can read and prompts it can reuse. It is the layer that turns your managed APIs and integration flows into something an agent can use.
The 2026-07-28 revision was the biggest change since authorization was added. The protocol core is now stateless: the initialize handshake and the Mcp-Session-Id header are gone, every request carries its protocol version and client capabilities in _meta, and servers needing more input mid-call use Multi Round-Trip Requests instead of calling back into the client. In practice an MCP server now scales behind an ordinary HTTP load balancer, which is exactly what enterprise gateways expect. We cover the details in What MCP 2026-07-28 changes for integration architects.
Adoption is no longer in question. The Agentic AI Foundation reported over 110 million monthly SDK downloads in April 2026, every major cloud ships an MCP gateway, and in July 2026 it announced the first official credential, the Model Context Protocol Associate (MCPA), whose exam opened for enrollment in September 2026.
A2A: the agent-to-agent layer
A2A solves a different problem. When your procurement agent needs a supplier's availability agent to check stock, neither side should expose its internal tools, prompts or memory. A2A lets agents treat each other as opaque services: one agent publishes an Agent Card describing its skills and endpoints, the other sends it a message, and the work is tracked as a task with a defined lifecycle (submitted, working, input-required, auth-required, completed, failed, canceled, rejected).
A2A v1.0 defines three equivalent bindings (JSON-RPC, gRPC and HTTP+JSON), signed Agent Cards using JWS over a canonicalized card, multi-tenancy, and three ways to get results: polling, streaming and push notifications to a webhook. By its one-year mark in April 2026 the project reported more than 150 supporting organizations and built-in support in Azure AI Foundry, Copilot Studio and Amazon Bedrock AgentCore. Its technical steering committee includes AWS, Cisco, Google, IBM Research, Microsoft, Salesforce, SAP and ServiceNow. See A2A for enterprise integration.
MCP vs A2A: how to choose
| Question | MCP | A2A |
|---|---|---|
| What is on the other end? | A tool, API, dataset or prompt library | Another agent with its own reasoning |
| Who decides how the work is done? | The calling agent; the tool executes a defined function | The remote agent; the caller only states the goal |
| Typical duration | Milliseconds to seconds; one request, one result | Seconds to days; a task with status updates |
| What is shared? | Tool schemas and results | Messages and artifacts; internals stay hidden |
| Discovery | tools/list on a known server, often via a gateway or registry | Agent Card, typically at a well-known URL |
| Integration analogy | A synchronous API call through an API gateway | A B2B partner process exchanging business documents |
They are complementary, not rivals. A common enterprise pattern: an orchestrating agent delegates "prepare the quarter-end accrual" to a finance agent over A2A, and the finance agent uses MCP tools to read open purchase orders and post the journal. Both hops pass through governed gateways.
The commerce layer: UCP, ACP and AP2
If your agents never buy or sell, you can skip this section. If you run order-to-cash or procure-to-pay, it matters, because agent-placed orders will land in your ERP.
- UCP covers discovery, catalog, cart and checkout between a shopping agent and a merchant. Google positions it as compatible with A2A, AP2 and MCP rather than replacing them.
- ACP standardizes agent-mediated checkout inside an AI surface, with a delegated payment token so the agent never sees the card.
- AP2 answers the question that chargeback and audit teams care about: can we prove the human actually authorized this payment? It does this with signed mandates that travel with the transaction.
For integration teams the implications are concrete: orders need an "agent-originated" flag and the mandate or token reference stored with the sales order, reconciliation must tolerate agent retries (idempotency keys again), and fraud rules need to recognize agent traffic rather than block it.
Who governs what
The Agentic AI Foundation (AAIF) was formed under the Linux Foundation on 9 December 2025, with Anthropic contributing MCP, Block contributing goose and OpenAI contributing AGENTS.md. Solo.io's open-source agentgateway joined in June 2026 and A2A in August 2026. One neutral home for both horizontal protocols reduces the risk that either fragments or is steered by one vendor, which is the main objection enterprise architecture boards raised in 2025.
Standards bodies are now working one layer down, on identity. NIST launched its AI Agent Standards Initiative in February 2026, explicitly treating MCP and A2A as interoperability baselines, and the IETF has several drafts on agent identity and delegated authorization. See Agent identity and authorization.
What this means for an integration landscape
- Put both protocols behind a gateway. MCP and A2A traffic are new kinds of API traffic. They need the same authentication, quotas, logging and kill switches as everything else. Several gateways (agentgateway, Kong AI Gateway, IBM ContextForge, AWS AgentCore Gateway) already proxy both.
- Keep MCP servers thin. Expose existing managed APIs and integration flows as tools; do not build new back doors to systems of record.
- Treat Agent Cards like partner profiles. Onboard external agents the way you onboard B2B partners: verify the signed card, agree scopes, test in a sandbox, monitor.
- Design for long-running work. A2A tasks and MCP's tasks extension can run for hours. Correlation IDs, compensation and idempotent writes from classic integration orchestration apply directly.
- Version-stamp everything. Record which MCP revision and A2A version each server and agent supports. Both specifications have formal deprecation policies now; MCP's guarantees at least twelve months' notice.
Frequently asked questions
Is A2A a replacement for MCP?
No. MCP connects an agent to tools and data; A2A connects an agent to another agent that keeps its internals private. Most enterprise designs use both, often in the same workflow.
Which agent protocol should an SAP landscape start with?
MCP. SAP Integration Suite's MCP Gateway exposes existing APIs and integration flows as governed MCP tools, which is the quickest safe route for agents to reach SAP data. Add A2A when agents in different platforms, such as Joule and a non-SAP agent, need to delegate work to each other.
What happened to IBM's Agent Communication Protocol (ACP)?
It merged into A2A under the Linux Foundation in August 2025 and is no longer developed separately. The ACP most people mean in 2026 is the Agentic Commerce Protocol from OpenAI and Stripe.
Do I need UCP, ACP or AP2?
Only if agents will buy or sell on someone's behalf. If they will, plan for agent-originated orders in ERP: store the mandate or payment token reference, enforce idempotency and adjust fraud and reconciliation rules.
Sources and further reading
- Model Context Protocol: 2026-07-28 key changes
- A2A: What's new in v1.0
- Linux Foundation: formation of the Agentic AI Foundation (Dec 2025)
- Linux Foundation: A2A passes 150 organizations (Apr 2026)
- Forbes: Agent2Agent joins the Agentic AI Foundation (Aug 2026)
- AAIF: Introducing the MCPA certification (Jul 2026)
- AAIF: MCPA exam opens for enrollment (Sep 2026)
- Crystl: A2A vs ACP vs MCP