On April 22, 2026, Google and OpenAI both launched enterprise agent platforms within hours of each other. The race to control how companies automate work just went from theoretical to procurement-table urgency. Within a single news cycle, Alphabet announced the Gemini Enterprise Agent Platform at Google Cloud Next ’26, while OpenAI rolled out ChatGPT workspace agents for its Business, Enterprise, and Education tiers. Two of the largest AI vendors in the world decided the same morning was the right moment to declare that autonomous agents are now enterprise infrastructure, not a research curiosity.
The launch is more than a coincidence. It signals that the next phase of AI competition is no longer about who has the best model — frontier models from Anthropic, OpenAI, and Google are within a few percentage points of each other on every major benchmark. The competitive surface has moved up the stack to orchestration, governance, cross-vendor interoperability, and the hardware economics that determine whether running millions of agents is even affordable. This is the technical breakdown of what each vendor shipped, what the open protocols underneath them do, what it costs, and what enterprise teams should actually do about it.
What Google Actually Launched on April 22
The Gemini Enterprise Agent Platform is the evolution of Vertex AI, not a replacement. According to the official Google Cloud launch post, it bundles five distinct components that previously had to be stitched together by hand. The Agent Development Kit (ADK) is the code-first, model-agnostic framework for building agents in Python, with a new graph-based architecture for orchestrating sub-agents. Agent Studio is the low-code visual builder for teams that prefer drag-and-drop workflows. Agent Runtime, formerly Agent Engine, is the deployment layer that hosts and executes agents at scale. Agent Registry is the catalog where every production agent lives with its version history and dependencies. Agent Gateway is the centralized security and policy layer that enforces who can invoke which agent, with what permissions, and what data can flow.
The platform is deliberately model-agnostic. It supports the Gemini 3.1 series as the first-party default, but it also accepts Anthropic Claude as a drop-in alternative for teams already standardized on that family. In Sundar Pichai’s Cloud Next ’26 keynote recap, he framed the launch as the moment the industry stopped asking how to build a single agent and started asking how to manage thousands. The components reflect that pivot. Building one agent is a developer problem; governing ten thousand of them across a regulated enterprise is an operations problem, and the platform is structured accordingly.
The other half of the announcement — two new chips designed specifically for the “agentic era” — signals how seriously Google is taking the cost economics of running agents at scale. That part deserves its own section below.
Agent2Agent: The Open Protocol That Makes Cross-Vendor Agents Possible
Underneath every multi-agent system, regardless of vendor, you eventually need a way for agents built by different teams, on different frameworks, running on different clouds, to talk to each other. The Agent2Agent (A2A) protocol is Google’s answer, and unlike most vendor-led standards, it has been donated to the Linux Foundation for vendor-neutral governance. The specification lives in the a2aproject/A2A repository on GitHub, defined in Protocol Buffers under specification/a2a.proto. The protocol exposes a small set of HTTP endpoints: POST /message:send for request-response communication, POST /message:stream for streaming, and GET /tasks/{id=*} for task lifecycle management.
What makes A2A architecturally interesting is that it is decentralized. There is no central router or registry that all agents must register with. An agent publishes a capability card describing what it can do; another agent discovers that card via standard discovery mechanisms and invokes it directly. The protocol does not assume any specific transport, authentication scheme, or data serialization format beyond Protocol Buffers. This matters because enterprise deployments rarely have the luxury of mandating a single framework — a Fortune 500 procurement team will end up running agents built in Python, TypeScript, and Go, hosted on GCP, AWS, and Azure, and they need them to interoperate without writing custom glue for every pairing.
The Linux Foundation move was the strategic turning point. As of June 2025, more than 50 companies had committed to the project, including Google, AWS, Cisco, IBM, Microsoft, Salesforce, and SAP. Cross-vendor commitment at that scale is rare for an AI protocol, and it is the main reason enterprise architects should treat A2A as a serious planning assumption rather than a Google-only initiative. If you are sketching a multi-agent architecture today, designing around A2A from day one is the lowest-regret choice.
OpenAI Fired Back the Same Day With Workspace Agents
OpenAI did not let Google’s announcement stand uncontested. According to The Verge’s same-day coverage, OpenAI rolled out ChatGPT workspace agents for Business, Enterprise, and Edu plans within hours of Google’s launch. The official OpenAI documentation describes them as persistent, cloud-based agents that automate multi-step workflows across third-party applications including Slack, Salesforce, and Google Drive. Admins control access via role-based governance; users can build, share, and schedule autonomous agents that maintain organizational memory.
The strategic positioning is different from Google’s. Where Google is selling to enterprise architects who already use Vertex AI, OpenAI is selling to the line-of-business buyer who already pays for ChatGPT seats. The underlying toolkit — the Agents SDK and Responses API — is the developer-facing surface, while workspace agents are the end-user-visible product layer. The Responses API is the evolution of Chat Completions: a stateful, unified interface for either single-call or multi-step agentic loops without requiring the Agents SDK’s full orchestration overhead. The Assistants API, by contrast, is being sunset on August 26, 2026 — a useful detail for any team still maintaining code against the older interface.
For enterprise teams evaluating both platforms, the practical question is not “which is better” but “which maps to your buying motion.” If your procurement cycle goes through a central platform team with a Vertex AI budget already approved, the Gemini Enterprise Agent Platform is the lower-friction path. If your line-of-business leaders already have ChatGPT Enterprise licenses and are pushing from the bottom up, workspace agents are where you will see faster time-to-value, particularly for the Salesforce/Slack/Google Drive integration patterns OpenAI has prioritized. The two approaches are not mutually exclusive — A2A is designed precisely to let agents built on either side of this divide talk to each other.
The MCP Question: Does Model Context Protocol Make A2A Redundant?
One of the more common architectural confusions in 2026 is conflating Anthropic’s Model Context Protocol (MCP) with Google’s A2A. They solve different problems and are designed to coexist. MCP is the protocol between an AI agent and the tools or data sources it needs to do its job — the agent-to-integration layer. A2A is the protocol between two independent AI agents that need to coordinate — the agent-to-agent coordination layer. An agent typically uses MCP to read from Salesforce or query a database, and uses A2A to negotiate a task handoff with another agent.
The MCP specification itself moved significantly in 2026. According to the 2026-07-28 specification release notes, the protocol transitioned from a stateful, handshake-based architecture to a stateless request-response model. Modern clients declare their protocol version in the _meta field of every request; legacy clients using the older initialize handshake still work via server-side era detection. The migration was driven by horizontal-scaling requirements: stateless request-response is dramatically easier to load-balance across thousands of agent instances than the previous stateful model. The versioning specification lists five architecture eras now in play — pre-2025, 2025-06-18, 2025-11-25, 2026-06-16, and 2026-07-28 — and any production deployment needs to declare which era it targets.
For most enterprises, the practical answer is to support both. Google’s documentation explicitly positions A2A and MCP as complementary. The agent economy that emerges over the next 18 months will not have a single protocol; it will have a small number of interoperable standards, each handling a different layer. Building today means designing for the coexistence: A2A on the agent-to-agent boundary, MCP on the agent-to-tool boundary, and either the Agents SDK or ADK on the agent-to-orchestration boundary depending on vendor choice.
The Hardware Behind the Agent Economy: TPU 8t and TPU 8i
Google did not just ship software on April 22. The TPU 8t and TPU 8i technical deep dive lays out two specialized eighth-generation TPUs, each designed for a different half of the agent workload. TPU 8t is the training chip: 9,600 chips per superpod, 3D torus network topology, SparseCore accelerators for embedding-heavy workloads, and the Virgo fabric that scales near-linearly beyond one million chips. TPU 8i is the inference and reinforcement-learning chip: lower precision, optimized for the latency-sensitive serving pattern that agent economies demand.
The economic argument is the point. Running a single agent in a pilot costs pennies. Running ten thousand agents continuously, each making dozens of model calls per minute, is a fundamentally different cost structure. As Pichai framed it in his keynote recap, the bet is that “managing thousands of agents cost-effectively” requires inference silicon that does not yet exist in the general-purpose GPU fleet. By vertically integrating the chip design with the agent platform, Google is making a long-term bet that infrastructure lock-in will follow from agent-platform adoption, not the other way around.
For enterprises, the practical implication is a shift in the build-vs-buy calculus. If your current AI infrastructure is GPU-heavy (Nvidia H100/H200, AWS Trainium, AMD MI300X), you have a choice: stay on the general-purpose ecosystem and pay the inference premium, or move workloads to GCP and benefit from the vertically integrated TPU economics. The TPU path saves money at scale but introduces migration friction. The decision becomes concrete when your agent fleet crosses roughly 1,000 concurrent agents in production — the cost gap at that scale is large enough to justify a multi-quarter migration.
Real-World Adoption: The HSBC Case Study
The most concrete validation of the agent-platform thesis came two months later, on June 17, 2026, when HSBC and Google Cloud announced a multi-year strategic partnership anchored on the Gemini Enterprise Agent Platform. The deal covers three flagship areas: hyper-personalized wealth management support for relationship managers, enhanced financial crime risk management, and AI-powered tools for frontline staff across HSBC’s global operations. HSBC expects the partnership to enable more than 200 AI use cases with an estimated total value exceeding US$100 million. Bloomberg’s subsequent reporting confirmed the dollar figure.
HSBC matters as a reference because it is a heavily regulated, multi-jurisdiction enterprise — exactly the customer profile that skeptics point to when they argue agents are not ready for production deployment. The fact that a tier-one global bank signed on with a quantified ROI expectation is more meaningful than another vendor pilot. It is the first large-scale data point that enterprise agents can clear the bar of regulated-industry compliance, and the specific use cases (financial crime, wealth management) are exactly the ones where false-positive costs are highest. If agents can deliver measurable value in that environment, the regulatory-bar argument for agent adoption weakens substantially.
The partnership is also a competitive signal. Anthropic’s Claude family was named as a supported model inside Gemini Enterprise Agent Platform at launch; HSBC’s deal validates that mix. If the partnership succeeds, expect Anthropic’s other financial-services customers to evaluate the same architecture, and expect OpenAI to feel pressure to ship a comparable flagship enterprise reference customer with quantified ROI.
What It Costs: Vertex AI Agent Engine Pricing Breakdown
Per the pricing section of Google’s launch post, Vertex AI Agent Engine uses a pay-as-you-go model with no flat monthly fee. Four additive usage meters determine the bill. Agent Engine Runtime is $0.0864 per vCPU-hour and $0.0090 per GB-hour of memory, metered per second with idle time excluded. Sessions and Memory Bank operations are $0.25 per 1,000 events. Vertex AI Search costs $1.50 to $6.00 per 1,000 queries depending on tier (Standard, Enterprise, or Conversational). Foundation model usage (Gemini 3.1, Anthropic Claude, etc.) is billed separately at standard model rates. New accounts get 50 vCPU-hours and 100 GiB-hours free per month.
Translate that into enterprise budgets. A single production agent running on a 2-vCPU instance with 8 GB of memory, online 24/7, costs roughly $145 per month in compute ($0.0864 × 2 vCPU × 730 hours) plus memory ($0.0090 × 8 GB × 730 hours = $53). Add 100,000 session events per month ($25) and foundation-model calls (typically $500-$5,000 depending on call volume and model choice), and you are at $700-$5,500 per agent per month. Multiply by 1,000 production agents and you are looking at a seven-figure annual bill. The pilot-to-production conversion math has to be done in dollars before it is done in capabilities.
The free tier matters for pilots but disappears fast. Fifty vCPU-hours per month is enough to run a handful of agents for evaluation but not enough to support anything resembling production. The realistic onboarding path is to budget $5,000-$50,000 for the first three months of agent-platform experimentation, then size production costs against actual usage rather than projection.
The Realistic Enterprise Outlook: 40% of Apps Will Have Agents by End of 2026
Gartner’s forecast from August 2025 projected that 40% of enterprise applications would feature task-specific AI agents by the end of 2026, up from less than 5% in 2025 — an eightfold adoption jump in 18 months. Spending on purpose-built AI agent software was forecast to reach $206.5 billion in 2026, a 139% increase over 2025. The forecast is now roughly half-realized: enterprise intent to adopt is high (over 60% of organizations experimenting per third-party surveys), but the production deployment rate is closer to 25%.
The pilot-to-production gap is the defining metric of 2026. Most enterprises have at least one agent in a pilot phase; far fewer have agents in production handling regulated workflows with measurable ROI. The gap exists for predictable reasons: governance frameworks are still being written, integration with legacy systems is harder than vendor demos suggest, and the observability tooling for multi-agent systems is years behind the observability tooling for traditional distributed systems. The platforms Google and OpenAI launched in April will not close the gap by themselves — they close the build-and-deploy half of the gap, but the run-and-govern half is still open.
According to Axios’s same-day coverage, the strategic intent behind the Google launch was specifically to address that gap — to make the production half of the journey easier than the build half. Whether that intent translates to measurable outcomes over the next two quarters is the question every enterprise architect should be tracking. The platforms are ready; the runbooks are not.
What to Do About It
If you are an enterprise architect evaluating agent platforms today, the move is not to wait for the dust to settle. The dust will not settle — A2A and MCP will coexist, the model layer will continue to commoditize, and the differentiation will remain at the orchestration and governance layers where vendor lock-in is most defensible. Pilot one production workflow against both Gemini Enterprise Agent Platform and OpenAI Agents SDK within 30 days. Use A2A for any cross-vendor coordination you cannot avoid. Track three metrics over the next two quarters: protocol convergence (will MCP and A2A merge or stay distinct?), hardware lock-in (does TPU 8i economics pull workloads off your existing GPU estate?), and your own pilot-to-production conversion rate against the Gartner 25% benchmark.
The platforms shipped on April 22 are infrastructure, not experiments. The companies that treat them as such — piloting fast, measuring honestly, cutting what does not convert — will be the ones that own the next layer of the enterprise AI stack. Everyone else will spend 2027 explaining why their agent pilot never made it to production.
Frequently Asked Questions
What is the Gemini Enterprise Agent Platform?
It is Google’s unified enterprise agent platform launched on April 22, 2026. It bundles the Agent Development Kit (code-first, model-agnostic Python framework), Agent Studio (low-code visual builder), Agent Runtime (formerly Agent Engine, the deployment layer), Agent Registry (catalog with version history), and Agent Gateway (centralized security and policy enforcement). It evolved from Vertex AI and supports the Gemini 3.1 model series plus Anthropic Claude as drop-in alternatives.
How is Agent2Agent (A2A) different from MCP?
A2A and MCP solve different problems. MCP (Anthropic-led) is the protocol between an AI agent and its tools and data sources — the agent-to-integration layer. A2A (Google-led, now Linux Foundation) is the protocol between two independent AI agents — the agent-to-agent coordination layer. They are complementary: an agent uses MCP to access tools and A2A to negotiate with other agents. Adoption overlaps significantly; Google supports both.
Did OpenAI really launch workspace agents the same day as Google?
Yes. OpenAI announced ChatGPT workspace agents for Business, Enterprise, and Edu plans on April 22, 2026, within hours of Google’s Gemini Enterprise Agent Platform launch. Workspace agents automate multi-step workflows across Slack, Salesforce, and Google Drive, with admin controls for permissions and governance. The two announcements together mark the start of what is now called the “enterprise agent war.”
How much does the Vertex AI Agent Engine cost?
Agent Engine Runtime bills at $0.0864 per vCPU-hour and $0.0090 per GB-hour of memory, metered per second with idle time excluded. Sessions and Memory Bank events are $0.25 per 1,000 events. Vertex AI Search costs $1.50 to $6.00 per 1,000 queries depending on tier (Standard, Enterprise, Conversational). Foundation models are billed separately at standard rates. New accounts get 50 vCPU-hours and 100 GiB-hours free per month.
What is the practical difference between TPU 8t and TPU 8i?
TPU 8t is Google’s eighth-generation training TPU: 9,600 chips per superpod, 3D torus network topology, SparseCore accelerators for embedding-heavy workloads, and the Virgo fabric that scales near-linearly beyond one million chips. TPU 8i is the inference and reinforcement-learning chip, optimized for low-latency serving and RL workloads. Google launched both on April 22, 2026 because running millions of agents simultaneously needs cheaper inference than existing GPU silicon provides.