The last line of Vertex AI's documentation didn't announce a deprecation. There was no press release, no migration timeline, no apologetic blog post about a beloved product's sunset. At Google Cloud Next '26 on April 22, 2026, Google Cloud declared that all Vertex AI services and roadmap evolutions would henceforth be delivered exclusively through something called the Gemini Enterprise Agent Platform , and in doing so, signaled a complete break with how enterprise cloud computing has worked for the past decade.
What Actually Happened
On April 22, 2026, at the opening keynote of Google Cloud Next '26, Google introduced the Gemini Enterprise Agent Platform , a comprehensive, end-to-end workspace to build, scale, govern, and optimize AI agents. The announcement absorbed Vertex AI entirely: all future Vertex AI capability development now flows exclusively through the Agent Platform. The platform is live at console.cloud.google.com/agent-platform/overview, and Google has confirmed no new features will be added to the standalone Vertex AI product line.
The platform organizes enterprise AI across four pillars: Build, Scale, Govern, and Optimize. On the Build side, Agent Studio offers a no-code visual interface, while the Agent Development Kit (ADK) provides a code-first framework processing more than 6 trillion tokens monthly across Gemini models. The ADK introduces a graph-based framework for sub-agent networks and an Agent Sandbox , a hardened, isolated environment for testing agent behavior before production. The Agent Garden provides pre-built templates covering code modernization, financial analysis, and invoice processing. The Scale pillar delivers sub-second cold starts, agent provisioning in seconds, and an Agent Memory Bank for persistent long-term context with high-accuracy recall through Memory Profiles.
Why This Matters More Than People Think
The Gemini Enterprise Agent Platform is not an upgrade to Vertex AI. It is a declaration that Google Cloud has decided enterprise AI is no longer about building models , it is about governing outcomes. Vertex AI was a toolbox for ML engineers who wanted Google's infrastructure to train and serve their own models. The Agent Platform assumes a fundamentally different customer: an enterprise that wants autonomous agents accomplishing business outcomes, not data scientists fine-tuning models. Google is not just repositioning a cloud product; it is repositioning its entire vision of what enterprises will pay for over the next decade.
Early customer evidence is already production-scale, not proof-of-concept. Comcast is deploying the platform to move "beyond scripted automation to conversational generative intelligence" in its Xfinity Assistant. L'Oréal is using it to shift "from workflow automation to autonomous, outcome-oriented orchestration" via Model Context Protocol integrations. PayPal has deployed the Agent Payment Protocol (AP2) for trusted agent-based commerce. Payhawk reported a 50%+ reduction in expense submission time after deploying the Memory Bank feature. These are production replacements for existing enterprise software workflows, not pilots.
The Competitive Landscape
Google made two moves at Cloud Next '26 that speak directly to competitive dynamics. First, the Gemini Enterprise Agent Platform exposes access to more than 200 models through Model Garden , including Anthropic's Claude Opus 4.7, Sonnet, and Haiku alongside Gemini 3.1 Pro and open models like Gemma 4. The fact that Google is hosting a competitor's flagship model inside its own enterprise AI platform is not charity. It is a calculated admission that winning the infrastructure and governance layer , not the model layer , is the real prize, and that offering the broadest model selection is how you lock in the governance budget.
Microsoft launched Agent 365 to general availability on May 1 at $15 per user per month , just nine days after Google's announcement. The timing was not coincidental. Both companies are now explicitly racing for the "agent governance layer": the system of record that IT departments use to discover, manage, secure, and audit every AI agent running in their organization. AWS has Bedrock Agents, but Google's Govern pillar , featuring cryptographic Agent Identity, Agent Gateway with Model Armor prompt-injection protection, real-time Agent Anomaly Detection, and Agent Threat Detection , goes further than any publicly disclosed competitor offering in operational agent security.
The most important thing Google did at Cloud Next '26 was not launch a product. It was write a retirement notice for an entire vision of enterprise cloud computing , and the industry barely noticed. Vertex AI represented the idea that the cloud's job was to give smart people fast access to powerful compute and models, then get out of the way. The Agent Platform represents the opposite: that the cloud's job is to become the cognitive infrastructure of the enterprise , knowing every agent, every task, every output, and governing all of it continuously. The power relationship between enterprise and cloud has fundamentally flipped.
This shift carries an implication most press coverage missed: the ML platform market as it existed from 2020 to 2025 is effectively over. Companies that built on Vertex AI as a training and serving substrate now face a choice , migrate to the Agent Platform's opinionated architecture, or move to a competitor still willing to sell lower-level abstractions. That choice will not be made by ML engineers. It will be made by CFOs and CISOs who see a governance platform promising audit trails, anomaly detection, and budgetary control over AI, versus a legacy ML platform offering none of those things.
There is a strategic signal embedded in the 6 trillion tokens-per-month figure that is easy to miss. This is not just a marketing number. It is a moat declaration. Google is processing more agent tokens than most enterprises can model, and it is using that scale to train its Agent Evaluation system , the continuous scoring and multi-turn autorater , on real-world agent behavior. Every enterprise that builds on the Agent Platform makes the platform more accurate for all others. This is the network-effect dynamic that made Gmail and Google Search impossible to compete with on equal footing, now applied to enterprise AI infrastructure. The platform improves with scale in ways no individual enterprise's internal AI system can replicate.
Perhaps most counterintuitively: the inclusion of Claude Opus 4.7 in Model Garden should concern Anthropic more than it helps them. Yes, Anthropic gains enterprise distribution through Google's channel. But every enterprise accessing Claude through Google's Agent Platform builds its observability, governance, and compliance workflows around Google's tooling , not Anthropic's. When that enterprise eventually decides which AI provider to standardize on at the infrastructure level, it is already standardized on Google. Distribution through a competitor's platform is a short-term revenue gain and a long-term dependency risk that Anthropic's enterprise sales team will feel acutely by 2027.
What to Watch Next
The most important indicator over the next 30 days is Vertex AI customer migration behavior. Google has not announced an official deprecation timeline, but the Cloud Next '26 message is unambiguous: no new capabilities are coming to standalone Vertex AI. Watch for changes in Vertex AI documentation redirecting users to Agent Platform equivalents, and for Google Cloud account team communications to enterprise Vertex AI customers about migration paths. The enterprises that engage first will have the most influence over how the migration tooling is designed , and the most leverage in pricing negotiations.
Over the next 90 to 180 days, the metric that matters most is the number of long-running agents deployed in production on the platform , agents that maintain state for multi-day periods within secure cloud sandboxes. This is the transition indicator from AI as a productivity tool to AI as a permanent operational layer. If production long-running agent counts grow faster than daily active agent counts, it signals that enterprises are not experimenting with the Agent Platform , they are structurally integrating it in ways that are very difficult to reverse. That is the adoption signal that will define the next wave of enterprise AI infrastructure investment decisions and the competitive positioning of every cloud provider in the governance layer race.
When Google decided that all of Vertex AI's future belongs to the Agent Platform, it wasn't a product decision , it was a declaration that the era of building enterprise AI is over, and the era of governing it has begun.
Key Takeaways
- Vertex AI absorbed entirely on April 22, 2026 , all future development flows exclusively through the Gemini Enterprise Agent Platform, with no new features coming to the standalone Vertex AI product
- 6+ trillion tokens processed monthly through ADK , with sub-second cold starts and Agent Memory Bank for persistent context; Payhawk reported 50%+ reduction in expense submission time in production
- 200+ models including Claude Opus 4.7 in Model Garden , Google hosts Anthropic's flagship directly on its platform, signaling that infrastructure governance ownership, not model superiority, is the real enterprise AI prize
- Deepest agent security stack publicly disclosed , cryptographic Agent Identity, Agent Gateway with Model Armor, real-time Agent Anomaly Detection, and Agent Threat Detection exceed any competitor's published offering
- Microsoft Agent 365 launched at $15/user/month nine days later , confirming the agent governance layer as the most contested new product category in enterprise cloud computing
Questions Worth Asking
- If Google is willing to host Claude Opus 4.7 on its own platform to win the governance layer, what does that tell you about where the actual value in enterprise AI is accumulating , and which companies are positioned to capture it?
- Enterprises that built ML infrastructure on Vertex AI now face a migration they didn't plan for , how does this change the risk calculus for betting on any single cloud AI platform going forward?
- The Agent Platform uses 6 trillion tokens monthly to continuously improve its evaluation systems , at what point does the platform's accumulated behavioral knowledge become a more defensible moat than any model it runs?