Google’s A2A Protocol and Workspace Studio: The Full-Stack Agentic Bet

Cloud Next as a Platform Declaration

Google Cloud Next 2026, which ran through the week of April 22, was less a conference than a platform declaration. Across the keynotes, Google laid down a coordinated set of moves: an evolved agent-to-agent protocol, a no-code agent builder inside Workspace, a general-purpose web-browsing agent, a designer canvas for custom workflows, and a growing garden of partner-built agents. Read together, these are not six separate product announcements — they are the components of one strategic thesis.

The thesis: enterprise value in the agent era will not accrue to a single vendor who owns a proprietary agent. It will accrue to the vendor who controls the model, the cloud infrastructure, the productivity suite, and the interoperability layer. Google is the only frontier provider with all four pieces already in-house, and it is betting that making them work well together — and work with competitors’ agents — is a stronger moat than locking down any single layer.

A2A v1.2 — Open, Cryptographic, In Production

The agent-to-agent (A2A) protocol, originally launched with more than 50 partners in 2025, is now at version 1.2 and, per Google’s numbers, running in production at 150 organizations. The protocol is governed by the Linux Foundation’s Agentic AI Foundation — a governance choice that signals openness and reduces the vendor-capture concern that otherwise blocks enterprise adoption.

Production deployments span the obvious list of major enterprise software providers: Microsoft, AWS, Salesforce, SAP, ServiceNow. The technical additions in v1.2 include cryptographic signature verification — each agent-to-agent exchange can now be signed and verified, which is table stakes for deploying multi-vendor agent networks in regulated industries.

The practical implication is significant. If an enterprise’s CRM agent (Salesforce), ERP agent (SAP), and ITSM agent (ServiceNow) can all hand off tasks to one another via a signed, audited protocol, a class of cross-system workflows that previously required custom integration code becomes a protocol-level capability. That is the long-term bet: interop as platform, not interop as integration project.

Workspace Studio — Agents Inside the Productivity Suite

Workspace Studio is Google’s answer to Microsoft Copilot Studio: a no-code agent builder for Gmail, Docs, Sheets, Drive, Meet, and Chat. Business users build automations with natural-language prompts rather than code, and the builder integrates with common third-party apps (Asana, Jira, Mailchimp, Salesforce) plus webhooks and custom Apps Script for advanced cases.

The rollout spans business, enterprise, and education Workspace customers — meaning the agent builder lands in front of hundreds of millions of existing Workspace seats without any additional procurement. This is the productivity-suite play and it is structurally different from OpenAI’s Frontier-platform approach. OpenAI asks enterprises to adopt a new platform; Google embeds the agent layer inside the suite employees already use.

Neither model is strictly better — but they will appeal to different enterprise profiles. Heavy Workspace shops will likely default to Studio. Firms already deeply invested in building custom agent platforms are more likely to stay with Frontier or Anthropic’s API.

The New Agent Products

The full product lineup announced at Cloud Next 2026 tells the completeness story:

• Project Mariner — a web-browsing agent scoring 83.5% on the WebVoyager benchmark, capable of handling 10 concurrent tasks. This is Google’s bet on the open-web research use case, where Gemini has historically been stronger than Claude.
• Agent Designer — a visual workflow canvas currently in preview, for engineers who want more control than Workspace Studio offers.
• Agent Engine Sessions and Memory Bank — persistent context for agents, now generally available. This addresses one of the hardest problems in production agents: giving an agent stable long-term memory across sessions without burning token budget re-introducing state.
• Agent Garden — a pre-built library of solutions for customer service, data analysis, and creative tasks.
• Partner agents — Box, Workday, Salesforce, ServiceNow, Dun & Bradstreet, and S&P Global agents are all in the garden, ready to be composed into workflows via A2A.

The Gemini 3.x model family underpins these products. Gemini 3 Pro and Flash, released in late 2025 and iterated through early 2026, deliver a 15% accuracy improvement over the previous generation (Gemini 3 Flash). Gemini 3.1 Pro is in preview for advanced reasoning, and Gemini 3.2 is expected next with an expanded context window beyond one million tokens. For agent workflows that need to reason over very long inputs — an entire codebase, a case file, an extended conversation history — Gemini’s context advantage is a genuine differentiator.

The Competitive Stack Up

Reading the April 2026 announcements side by side is instructive:

• Google: Full-stack bet. Model (Gemini 3.x) + infrastructure (GCP) + productivity suite (Workspace) + interop protocol (A2A) + partner agent garden.
• OpenAI: Platform-first bet. Frontier + Agents SDK + Codex + unified superapp thesis.
• Anthropic: Model-reliability bet. Claude Opus 4.7 with industry-leading SWE-bench and long-horizon agent scores, distributed through AWS Bedrock, Google Vertex AI, and Microsoft Foundry.
• Microsoft: Suite-integration bet. Copilot Studio embedded deep inside Microsoft 365.

Each vendor is betting on a different pivot axis for the agent economy. For enterprise buyers, the right frame is not “which vendor wins” but “which mix matches our stack and risk tolerance.”

Thomas Kurian’s quote at Cloud Next — “if you want to adopt a technology successfully, you need to pick a few important projects and do them well, rather than spraying on a lot of little projects” — is the honest advice underneath the full-stack pitch. Enterprises that deploy 50 small agent pilots in 2026 will not see returns. Those that pick three workflows, deploy serious agents, and measure end-to-end completion rates will build the institutional skill to operate agents at scale by 2027.

What Enterprises Should Do

Three practical steps:

First, evaluate A2A as interop insurance. Even organizations not immediately adopting Google agents should track A2A because it is the most credible multi-vendor agent protocol. If it continues to gain adoption, designing for A2A compatibility now prevents a re-platforming cost later.

Second, pilot Workspace Studio if Workspace is the primary suite. The marginal cost of enabling Studio inside an existing Workspace contract is small; the value of putting a no-code agent builder in front of knowledge workers is real.

Third, benchmark Project Mariner against current research agents. 83.5% on WebVoyager is competitive, and the 10 concurrent task handling is a genuine operational improvement. For open-web research workloads, Mariner belongs in the evaluation set alongside OpenAI’s deep-research agents and Anthropic’s browser-use agents.

What Enterprise Architects Should Do About the A2A Stack

The A2A protocol, Workspace Studio, and Project Mariner are three entry points to the same underlying architectural shift: the enterprise stack is moving from applications that humans operate to agents that operate applications. The right response is not to pick a vendor but to design for the transition. The following three steps address the specific decisions that enterprise architects face in the next two quarters.

1. Audit Your Current Vendor Integrations for A2A Compatibility Exposure

A2A v1.2 is already in production at 150 organizations including Microsoft, AWS, Salesforce, SAP, and ServiceNow. If your enterprise uses two or more of those platforms, you will encounter A2A-mediated agent hand-offs before you formally adopt any Google product. The practical move now is to audit your integration architecture: identify which of your existing software vendors have announced A2A support, map the workflows where cross-system task delegation already happens (CRM-to-ERP, ITSM-to-HR, contract-to-finance), and assess whether those workflows are likely to be agent-mediated within 24 months. Bloomberg’s coverage of Google Cloud Next 2026 quotes Kurian directly on the “few important projects” discipline — an audit of high-frequency cross-system workflows typically surfaces three to five candidates that account for 60 to 70 percent of integration maintenance burden and are therefore the highest-value A2A pilot targets.

2. Run a Workspace Studio Pilot on One Real Knowledge-Worker Workflow This Quarter

The marginal cost of enabling Workspace Studio for an existing Workspace business or enterprise subscriber is near zero — the builder ships inside existing licensing. The value test is whether knowledge workers in sales, operations, legal review, or finance can build automations that meaningfully reduce manual steps without requiring engineering intervention. Set a specific success criterion before enabling the pilot: a named workflow, a baseline step count, and a target reduction. Thomas Kurian’s Cloud Next guidance to focus on “three workflows done well rather than 50 pilots” is directly applicable here. A pilot that reduces a seven-step Gmail-to-Sheets workflow to two steps for 20 knowledge workers is a valid business case. A pilot that demonstrates Studio’s general capability without reducing any specific step count is not.

3. Design a Governance Framework for Multi-Vendor Agent Networks Before They Deploy

A2A v1.2’s cryptographic signature verification is the technical prerequisite for governed multi-agent operation: each hand-off can be signed, logged, and audited. But the governance framework — who authorises an agent to act on behalf of the enterprise, what decisions require human-in-the-loop confirmation, how audit logs are reviewed, and who is accountable when a multi-agent chain produces an error — must be designed by the enterprise, not the protocol. Anthropic’s model-reliability positioning and the EU AI Act’s requirements for high-risk AI systems both converge on the same point: production agent deployments in regulated industries need documented authorisation chains and escalation paths before they handle consequential decisions. Build the governance framework now, for the three candidate workflows identified in step one, before the first agent goes into production. Retrofitting governance to deployed agents is substantially more expensive and disruptive than designing it in from the start.

The Structural Lesson

The structural lesson of Google Cloud Next 2026 is not about any single product — it is about what kind of moat wins in the agent era. For decades, enterprise software moats came from data lock-in: once your CRM held five years of customer records, or your ERP held a decade of financial history, switching costs were prohibitive regardless of competitive alternatives. The A2A protocol is Google’s bet that the next generation of moats will come from interoperability instead — that the vendor who controls the protocol layer through which all agents communicate accumulates a structural position that is harder to challenge than any proprietary data silo.

That bet is historically unusual. Open protocols typically resist becoming moats because openness distributes the value across all participants. The Linux Foundation’s governance of A2A is designed precisely to prevent Google from capturing the protocol. But protocol governance and protocol leadership are different things: the vendor who writes the most production deployments, attracts the largest partner garden — Box, Workday, Salesforce, SAP, ServiceNow — and adds the most credible cryptographic and security features to each version maintains a de facto leadership position even without formal control. A2A’s 150 production organizations and v1.2 cryptographic signature verification are the early evidence that Google is executing this play effectively. The structural lesson for enterprise architects is not to avoid A2A — it is to participate in it with explicit abstraction layers that preserve optionality if the governance structure or the protocol leadership changes. Thomas Kurian’s Cloud Next advice to “pick a few important projects and do them well” is good tactical guidance; the structural context for that advice is that the projects you pick in 2026 will define the platform dependencies you live with in 2029.

Frequently Asked Questions

Sources & Further Reading

• Google Cloud Next 2026: AI agents, A2A protocol, Workspace Studio, and the full-stack bet — The Next Web
• Google releases new AI agents to challenge OpenAI and Anthropic — Bloomberg
• Agent2Agent (A2A) Protocol — Linux Foundation Agentic AI Foundation
• Google Cloud Next 2026 Keynote — Google Cloud Blog
• Fresh Agentic AI News and Developments — Crescendo