Agent protocols harden: MCP goes stateless, A2A passes 150 orgs
The 28 July 2026 MCP spec makes the protocol stateless and adds sandboxed MCP Apps; A2A reaches v1.2 with signed agent cards at the Linux Foundation
Summary
The agent-interoperability layer is standardising fast. Anthropic's Model Context Protocol, the de facto agent-to-tool standard, ~97M downloads, ships a 28 July 2026 specification that makes the protocol stateless via six enhancement proposals, adds MCP Apps (servers shipping interactive HTML rendered in a sandboxed iframe), and demotes Tasks to an extension. In parallel, Google's Agent2Agent protocol, now hosted by the Linux Foundation's Agentic AI Foundation, passed 150 organisations in production (Microsoft, AWS, Salesforce, SAP, ServiceNow) and reached v1.2 with cryptographically signed agent cards. The emerging consensus: MCP for vertical tool calls, A2A for horizontal agent-to-agent delegation, with a Q3 2026 joint spec as the first formal bridge.
By the numbers
- 2026-07-28, MCP specification release date (stateless protocol layer).
- 6, Specification Enhancement Proposals delivering statelessness.
- ~97M, cumulative MCP downloads.
- 150+, organisations running A2A in production.
- 12 months, minimum between MCP feature deprecation and removal.
Why it matters
These two specs decide how AI agents call tools and talk to each other across vendors, the plumbing of the agentic economy. Statelessness eases enterprise scaling; signed agent cards and OAuth 2.1 address the security gap that has blocked production. Whoever steers the specs steers the interoperability rules.
What to watch
- The 28 July 2026 MCP spec landing and MCP Apps' sandbox security record.
- The Q3 2026 MCP/A2A joint specification and shared authorization model.
- Whether OpenAI, Google and Microsoft keep converging or fork the stack.