A2A Adoption Forecast 2026–2028: Who's In, Who's Not

A2A adoption describes how strongly the Agent2Agent protocol is establishing itself as an open standard for communication between independent AI agents. As of 2026, A2A sits under the Linux Foundation and is backed by Google, Microsoft, AWS, Salesforce, SAP, ServiceNow and Cisco. This article situates the drivers, barriers and scenarios of adoption through 2028, explicitly as a forecast, not an established fact.

Important note upfront: Everything that goes beyond today's verifiable evidence base (adoption rates, market shares, the timing of production cross-vendor interoperability) is an assessment with substantial uncertainty bands. AI forecasts over a horizon of 18 to 24 months are fragile. Structural directions are more robust than specific milestones.

Quick answers

• Who's in? The major platform vendors: Google (initiator), Microsoft, AWS, Salesforce, SAP, ServiceNow and Cisco as founding members. In total, over 150 organizations support the standard (as of April 2026). This makes A2A the de facto standard for cross-platform agent collaboration.
• Who's not (yet) fully in? OpenAI is not a founding member. In August 2025, an OpenAI maintainer said there were no immediate plans to add A2A to the Agents SDK. OpenAI co-founded the Agentic AI Foundation and contributed AGENTS.md. Anthropic primarily drives MCP and approaches A2A via bridge patterns. Even founding members with strong proprietary ecosystems, such as Salesforce (Atlas) and Microsoft (Copilot Studio handoffs), still use A2A only at the outer edge and keep internal orchestration to themselves.
• What is the biggest risk? Fragmentation through proprietary AgentCard extensions. Not a vendor war, but creeping differentiation within the standard.

What A2A is and where it sits in the protocol stack

The Agent2Agent protocol was unveiled by Google at Cloud Next in April 2025 with over 50 partners and donated to the Linux Foundation on June 23, 2025, at the Open Source Summit North America in Denver. Google handed over the specification, the SDKs (Python, TypeScript) and the developer tooling. Founding members of the Linux Foundation project are AWS, Cisco, Google, Microsoft, Salesforce, SAP and ServiceNow.

Technically, A2A builds on JSON-RPC 2.0 over HTTPS, with streaming via Server-Sent Events and optional push notifications for long-running tasks. Each agent publishes an AgentCard, a JSON document with endpoint, capabilities, supported skills, modalities and authentication schemes. Tasks pass through a defined lifecycle (submitted → working → input-required → completed | failed | canceled). The decisive conceptual step: the internal logic of the remote agent remains opaque. A2A defines how agents talk to each other, not how they think. This allows a Salesforce agent to invoke an SAP agent without either exposing its prompts, models or memory to the other.

A2A does not stand alone. Together with MCP it forms a converging stack:

• MCP (Model Context Protocol, originally published by Anthropic in November 2024, donated to the Agentic AI Foundation under the Linux Foundation on December 9, 2025) governs agent-to-tool.
• A2A governs agent-to-agent.

The official reading, consistently advocated by Google, Salesforce and Microsoft, is: "MCP for capabilities, A2A for collaboration." Microsoft's own guidance puts it directly: MCP for tool and data access, Linux Foundation A2A for cross-platform agent-to-agent messaging.

The institutional drivers: why A2A has structural tailwind

The most important driver is Linux Foundation governance. It is the strongest conceivable structural protection against a vendor war and against unilateral control. IBM's competing ACP project was already merged into A2A on August 29, 2025, under LF AI & Data. The ACP team has ceased active development and contributed its technology to A2A. The result is a small, coherent open-protocol stack rather than a standards war.

The second driver is the anchoring in the enterprise platforms that DACH organizations already operate. Here A2A is mostly a product feature already (as of October 2026):

• Salesforce Agentforce 360 uses A2A as a cross-vendor handshake for delegation to third-party agents, while internal handoffs remain on the Atlas Reasoning Engine.
• SAP Joule is getting bidirectional A2A: third-party agents are meant to call Joule agents and vice versa. General availability of Joule's A2A capabilities is planned for Q4 2026. Agents built with the new Joule Studio, which SAP presented at Sapphire 2026, are to support MCP and A2A natively. According to SAP, Anthropic Claude is to become a primary reasoning and agentic capability in Joule.
• Microsoft Copilot Studio and Microsoft Foundry (formerly Azure AI Foundry) have integrated A2A.
• Google Gemini Enterprise, which absorbed Agentspace in October 2025, connects external agents via A2A.
• ServiceNow Now Assist uses A2A as well, and the AWS Bedrock AgentCore Runtime hosts A2A servers.

For DACH this is central: the SAP, Salesforce, Microsoft and ServiceNow estates are the entry points where companies first encounter A2A at all.

The barriers: where adoption realistically snags

Barrier 1: fragmentation through proprietary extensions. This is the most important open risk. Governance protects the core, but every founding member (Salesforce, SAP, Microsoft, Google, AWS, ServiceNow, Cisco) has commercial incentives to differentiate via AgentCard extensions, custom task states or "A2A-plus" capabilities. If AgentCards become vendor-specific, precisely the portability that justifies the protocol choice is lost. The watchlist signal for 2026 is therefore concrete: Does AgentCard remain a portable JSON document or does it become vendor-specific?

Barrier 2: competition from proprietary ecosystems. Many "multi-agent" deployments are in reality a single agent with many tool integrations. We call this a marketing multi-agent. Such systems do not need A2A at all. As long as internal orchestration (Salesforce Atlas, Copilot Studio handoffs, SAP Joule internal) remains platform-native and A2A is used only at the outer edge, the real depth of A2A usage is lower than the announcements suggest.

Barrier 3: unresolved governance and trust-boundary questions. Every A2A peer is opaque to the others. This is good for confidentiality, but difficult for audit trails, trust boundaries and, in the DACH context, data sovereignty. When a Salesforce agent invokes an SAP agent, which in turn triggers a Claude reasoning step via A2A, the data flow of a single task can cross several processor and residency boundaries. Under Art. 28 GDPR, this lengthens the processor chain. If a hop transfers personal data to a third country, Art. 44–49 GDPR apply as well: without an adequacy decision under Art. 45 for the destination, that hop needs its own transfer assessment. Hops within the EU and EEA do not trigger a transfer assessment. This slows production adoption in regulated industries.

Delineation: MCP vs. A2A, AGNTCY and NANDA

A common confusion is to read A2A as a competitor to MCP, AGNTCY or NANDA. That is wrong. The roles are clearly assigned:

• AGNTCY (Cisco, donated to the Linux Foundation on July 29, 2025) is not a competing protocol, but the identity, discovery and observability layer above an A2A mesh. A2A agents and MCP servers are discoverable via AGNTCY directories.
• NANDA (MIT Media Lab) is predominantly research-grade: a long-term vision for a decentralized agent identity, not a protocol used in production in DACH stacks (as of 2026).

Our assessment for architecture decisions in 2026: MCP + A2A is the converging industry stack. AGNTCY is the rational complement when cross-vendor identity matters. NANDA belongs on the watchlist, not on the critical path.

Vendor landscape: position and signal

The following table is a forecasting snapshot based on the documented signals (as of 2026), not a guaranteed roadmap.

Scenarios through 2028

Scenario A: convergence (the most likely direction). MCP + A2A establish themselves as a two-protocol world. In our view, a clean two-protocol state is the most likely status in 2027. AgentCard remains portable; AGNTCY takes over identity and observability where cross-vendor federation matters. Adoption grows, driven by the SAP, Salesforce and Microsoft estates.

Scenario B: quiet fragmentation. A2A formally remains the standard, but proprietary AgentCard extensions and "A2A-plus" features create de facto lock-in. Cross-vendor interoperability works in the demo, but not in production across multiple corporate estates. One open question decides between A and B: do SAP Joule, Agentforce and Copilot Studio deliver genuine cross-vendor A2A in production by the end of 2026, or does the number of production multi-agent deployments across the entire DAX 40 remain in single digits?

Scenario C: platform dominance instead of protocol. If one or two hyperscaler platforms make their internal orchestration so dominant that A2A remains only a peripheral function, the strategic significance of the open standard declines. Less likely given the broad founding coalition, but not ruled out.

A concrete example: A2A hop economics in a DACH corporation

An insurer operates, like Allianz with "Project Nemo" (seven specialized agents for food-spoilage claims following natural disasters, in production in Australia since July 2025, rolled out in under 100 days, with a documented 80% reduction in processing and settlement time for eligible food-spoilage claims under AUD 500, human-in-the-loop at payout), a hierarchical multi-agent architecture. Transposing this to a cross-vendor case:

```
Customer inquiry
→ Agentforce primary agent (Salesforce, internally Atlas)
→ A2A hop 1 → SAP Joule coverage agent (processor boundary + possibly cross-border)
→ A2A hop 2 → Claude reasoning step (processor boundary + possibly cross-border)
→ Audit agent (complete trace, one trace ID across all hops)
→ Human decides payout
```

Every A2A hop that passes personal data to another service provider adds a link to the data processing agreement chain. If a hop also goes to a third country without an adequacy decision, it needs its own transfer assessment. The example above has two hops: two additional links in the chain and, if both lead to such third countries, two transfer impact assessments. If all processors stay within the EU or EEA, the transfer assessment falls away, but the longer chain remains. This calculation is easy to miss in multi-agent design. The operational consequence: pin model versions in production flows, log all sub-agent prompts, tool calls and AgentCards used, and correlate them via a single trace ID.

For agencies and B2B decision-makers

For B2B decision-makers: treat A2A in 2026 as a likely, but not guaranteed, standard. The rational default sequence is to standardize MCP server policy, introduce A2A first within a single estate, publish AgentCards, and only then extend to a second estate. Anchor in procurement contracts the question: does the Agentforce, Joule or Copilot Studio contract guarantee A2A interoperability and MCP standards compliance, and what happens at exit? Hold back 15–25% of the AI budget as a trigger-based reserve for protocol and vendor events.

For marketing and AI agencies, A2A is the biggest product-strategy lever: publish an AgentCard for every productized agent, so that customer estates can invoke your agents directly from Agentforce, Joule or Copilot Studio without custom integration. Whoever builds A2A-discoverable from day one here will not be stuck behind a bespoke integration forever.

Blck Alpaca, based in Vienna, supports DACH companies with exactly this trade-off: sovereign, audit-proof multi-agent architectures based on MCP and A2A, with a clear separation of verified fact and forecasting assessment. Talk to us before committing to a tool 24 months in advance. The direction is set, the specifics are moving.