The announcement in one paragraph
At Google Cloud Next 2026 this week, Google confirmed that the Agent2Agent (A2A) protocol — the open spec for agents built on different platforms to call each other without custom integrations — is now running in production at 150 organizations, up from a pilot-heavy footprint a quarter ago. The protocol moved under the Linux Foundation's Agentic AI Foundation, shipped a 1.2 revision with cryptographically signed agent cards, and now has Microsoft, AWS, Salesforce, SAP, and ServiceNow running live A2A integrations. The broader keynote added a Vertex AI rebrand to Gemini Enterprise Agent Platform, new Workspace governance controls, and an MCP server for Drive, Gmail, and Calendar. The A2A line is the one worth reading twice.
Why "150 in production" is the number that matters
Every protocol announcement of the last three years has led with a partner logo count. A2A crossing into production at 150 organizations is a different claim from 150 companies signed a memorandum of interest. Production means someone's agent is calling someone else's agent right now, for a workload that breaks if it stops working. That is the threshold at which a protocol stops being a research exhibit and starts being a dependency.
It is also the threshold at which not speaking A2A becomes a procurement friction. An enterprise buyer looking at two agentic vendors — one that plugs into their existing ServiceNow or Salesforce agent fleet without custom glue, one that requires a six-week integration — will pick the first vendor every time, and the selection will increasingly be made before a technical bake-off ever happens.
What A2A actually standardizes, and what it doesn't
A2A is an interop spec, not a reasoning spec. It defines how agents describe their capabilities to each other, how they authenticate, how they pass tasks, and — in 1.2 — how they sign their identity cards so you know the agent you're talking to is actually the one it claims to be. It does not define what the agent does inside the black box, which model it uses, or how good it is at the task.
That matters because interop is a necessary but insufficient property. An agent can be fully A2A-compliant and still be bad at the thing you hired it for. The protocol's job is to keep the integration tax off the critical path, not to replace evaluation.
What it does to build-vs-buy
For three years the default answer for "should we build our own agent for X?" has been yes, because the available ones don't talk to our existing stack. A2A is quietly flipping that default. If the candidate vendors speak the protocol, and your CRM, ITSM, and ERP suite already do, the integration cost drops to near-zero — and the build case has to justify itself on quality and control alone, without the integration-pain tax on the buy side of the ledger.
A lot of half-built internal agent prototypes are going to get archived in Q2 as a result. The teams that built them did so to avoid integration pain that no longer exists.
Where we'd push back on the narrative
Two honest gaps worth naming.
The governance story is better than the security story. Moving to Linux Foundation governance is the right move and exactly the signal enterprise procurement is trained to look for. But A2A's threat model is still maturing — cross-agent prompt injection, delegated-auth scope creep, and reasoning-chain poisoning between trusted agents are live research problems, not solved ones. Google's own Workspace governance announcements this week implicitly acknowledge the gap: the agent inbox, audit controls, and prompt-injection defenses are retrofit perimeter work around a protocol that didn't ship with them built in.
"150 in production" is a number Google can't independently verify. It is a self-reported figure from vendors with an interest in signaling traction. Not a reason to discount it — the major names (Microsoft, AWS, SAP, Salesforce, ServiceNow) are all shipping integrations you can inspect — but a reason to weight it as a leading indicator, not a finished state.
The enterprise procurement angle
If you are running or advising on an AI agent procurement this quarter, three questions to add to every RFP:
- Does this agent expose an A2A interface, or will we need a bridge layer?
- Is the agent's identity card signed, and against which domain?
- What is the governance posture when this agent hands off to a third-party agent we did not select?
Vendors who can answer these cleanly are telling you they have read the spec. Vendors who cannot are telling you they will be behind on this in six months, and your integration budget will pay for the gap.
What we would do with this today
- Inventory the agents already in your stack. Most orgs have more than they think — sales ops agents in Salesforce, IT agents in ServiceNow, HR agents in Workday. The A2A question is whether those agents can cooperate without you writing glue.
- Build the evaluation harness now. The interop cost disappearing does not make the quality cost disappear. A2A-compliant agents still need benchmarking against the task, and the teams with evals in place will pick better vendors faster.
- Don't over-rotate. A2A solves a real integration problem but does not change which model is good at your workload. If the best agent for your task is from a vendor not yet shipping A2A, the right move might still be to buy it and bridge — just with an honest line item on the integration cost.
- Treat bridge-layer vendors with care. A cottage industry will emerge selling "A2A-compatibility shims" for agents that do not natively speak the protocol. Some of these will be useful for a quarter and dead weight by year-end. Contract accordingly.
The broader read
The pattern that keeps repeating in enterprise AI is that standards only get adopted after the interest bloc is wide enough that no single vendor can refuse. A2A now has that bloc. The competitive question for the next two quarters is not will this protocol stick — it will — but which vendors' agents get composed into which workflows once composition stops being expensive. The agents that win composition will win market share, and the agents that don't will get routed around.
The era of agentic AI where every vendor was an island is ending. The era where the best-composed fleet wins is starting. Procurement will feel this first; the rest of the stack will catch up.
