Google made a significant architectural declaration at its Cloud Next conference in Las Vegas on April 22. The company announced the Gemini Enterprise Agent Platform, a comprehensive platform for building, scaling, governing and optimizing AI agents, and simultaneously retired Vertex AI as a standalone brand. Going forward, all Vertex AI services and roadmap updates will be delivered exclusively through Agent Platform. The move signals that Google considers the era of isolated model playgrounds over, and that the enterprise technology stack should now be organized around agents as the primary unit of work.

For technology leaders evaluating their AI infrastructure, the practical question is not whether agents matter but what Google has actually built, how the pieces fit together and what the constraints are before committing to the platform.

What Agent Platform Is and Where It Comes From

Vertex AI, launched in 2021, gave enterprises a managed platform for training, tuning and deploying AI models. Agent Platform is its evolution, not its replacement. Existing Vertex AI APIs remain backward compatible, and organizations that built on Vertex AI can access the new brand directly inside their Google Cloud console without migration. To use the new capabilities, including Agent Runtime, Memory Bank and Agent Registry, enterprises need to enable the Agent Platform plan.

The platform is organized around four pillars. Build covers agent development tools. Scale addresses production runtime and persistent memory. Govern provides identity, registry and policy enforcement. Optimize handles testing, evaluation and observability. Model Garden provides access to more than 200 models, including Google's Gemini 3.1 Pro, Gemma 4 open-weight models and third-party models including Claude Opus, Sonnet and Haiku from Anthropic. The inclusion of competing models reflects a deliberate positioning. Google is betting that enterprises will choose the platform for its infrastructure and governance capabilities, not solely for its own models.

The Build Layer: Two Paths to Agent Development

Agent Studio is a generally available, low-code visual canvas for designing agent reasoning loops, setting triggers and building schedule-based automation without writing code. Agent Designer, the companion tool inside the Gemini Enterprise app, extends this to knowledge workers who want to build agents within their daily workflow.

The Agent Development Kit, widely known as ADK, serves code-first teams. It is a modular, model-agnostic framework that now includes a graph-based orchestration layer for defining structured logic across multiple cooperating agents. Google reports that more than six trillion tokens are processed monthly through ADK. The framework supports Python, Java and Go and is compatible with third-party orchestration frameworks, so teams are not required to abandon existing tooling. Agent Garden supplements both paths with prebuilt templates for use cases including financial analysis, code modernization and invoice processing.

The Scale Layer: Agents That Run for Days

Earlier generations of AI agents were built around short, discrete interactions. Agent Platform's scale layer is built for a different model, one where agents persist, accumulate context and execute tasks over extended periods.

Agent Runtime, now generally available, delivers sub-second cold starts and supports long-running agents that maintain state for multiple days, enabling use cases such as multi-day sales prospecting sequences and extended data pipelines that require human approval at defined stages. Memory Bank provides persistent context storage across sessions, allowing agents to carry forward information from previous interactions without requiring users to restate it each time.

The scale layer is already in production at several large enterprises. GE Appliances is running more than 800 AI agents across manufacturing, logistics and supply chain. Highmark Health's Sidekick AI assistant delivered $27.9 million in value in 2025 by automating research protocols for internal teams. Comcast rebuilt its Xfinity Assistant using ADK, shifting from scripted automation to conversational troubleshooting that resolves customer issues on first contact more consistently than its predecessor.

The Govern Layer: The Part Most Platforms Skip

The governance stack is where Agent Platform makes its most distinct architectural argument.

Agent Identity assigns every agent a unique cryptographic credential based on the SPIFFE standard, the identity framework used in zero-trust security architectures. Each agent receives an X.509 certificate and a SPIFFE ID tied to its lifecycle. Every action the agent takes is signed against this identity and logged, creating a verifiable audit trail mapped to defined authorization policies. The Agent Identity auth manager, which handles credential storage and delegation, is currently in preview.

Agent Registry provides a centralized catalog of every agent, tool and Model Context Protocol server across the organization. Only assets registered here are discoverable and available to users, giving IT a single control point to prevent agent sprawl.

Agent Gateway acts as the policy enforcement layer for all agent-to-agent and agent-to-tool traffic. It understands MCP and Agent-to-Agent protocols natively, inspects interactions in real time and applies Model Armor to guard against prompt injection, tool poisoning and sensitive data leakage. Agent Gateway's integration with Agent Runtime is currently in preview.

The governance approach differs from how AWS and Microsoft have addressed the same problem. AWS AgentCore , which reached general availability in October 2025, uses Cedar-language policy gateways to separate policy logic from agent code, with every policy decision logged through CloudWatch. Microsoft Foundry , generally available since Build 2025, integrates agents directly into the Entra ID identity system, giving every agent a dedicated Entra ID identity governed through the same administrative tools used for human users. Google's approach uses SPIFFE-based cryptographic identity combined with semantic governance policies that govern MCP server traffic at the content level, not just at the access control level.

Testing, Observability and Limitations

The optimize layer addresses a gap common in early enterprise agent deployments. Agent Simulation generates synthetic test scenarios to stress-test agent logic before production. Agent Evaluation provides systematic quality assessment through automated raters. Agent Observability delivers full execution traces and a real-time view into agent reasoning for performance monitoring and debugging.

Several constraints apply as of the April 22 launch. Agent Identity's auth manager and Agent Gateway's integration with Agent Runtime are both in preview. Enterprises should confirm general availability timelines before building production governance commitments around either. Agent Gateway is regional in scope, requiring deliberate architecture planning for multi-region deployments. Google has not published a unified pricing sheet. The billing model combines pay-per-token charges for model access, per-runtime-minute pricing for Agent Runtime and flat subscription pricing for Agent Registry and Gateway at the enterprise level.

The governance capabilities are also the primary source of vendor dependency. Agents relying on Memory Bank and Agent Registry lose those capabilities when moved outside Google Cloud. The lock-in sits at the runtime and governance layer, not at the model layer where Google explicitly supports third-party models.

What Technology Leaders Should Consider

The right platform depends on where an organization's existing infrastructure already lives. Enterprises running on Google Cloud with BigQuery, Google Kubernetes Engine and Workspace get the most native integration path. AWS-first organizations have a production-ready alternative in AgentCore. Microsoft-centric organizations benefit most from Foundry's Entra ID integration, which governs agents through the same tools used for human user management.

For any organization beginning an agentic AI program, the governance architecture decision should be made at the start of platform selection. Applying cryptographic identity and gateway policies to agents already running without controls is a re-engineering effort that compounds in cost with every additional agent deployed.