One platform to build, tune and run agents

At its Cloud Next conference, Google positioned the new Agent Platform as the successor to Vertex AI for enterprises that want to deploy autonomous or semi-autonomous agents at scale. Thomas Kurian, CEO of Google Cloud, framed the change as a consolidation: the model selection, model-building and tuning services teams used in Vertex AI now flow through the Agent Platform, Google said.

The platform ships with access to more than 200 models, including:

  • Google’s Gemini 3.1 Pro
  • Nano Banana 2
  • The open Gemma family
  • Third-party offerings such as Anthropic’s Opus 4.7

By letting organizations tap different models from a single control plane, Google aims to simplify model selection and matching a model to a task.

Developers can design agents end-to-end inside the platform: building agent logic, selecting models, tuning prompts and embedding recall. Google also added an Agent Development Kit that lets engineers break complex workflows into sub-networks, or sub-agents, so a single package can cover multi-step tasks that previously required manual handoffs or custom glue code.

The modular approach improves reasoning across large jobs. Agents can pass work to specialized sub-agents, get faster runtime for critical pieces, and keep relevant context longer through an integrated Memory Bank. The Memory Bank provides persistent context so agents retain a conversation’s thread across repeated calls.

Security, identity and pre-launch testing

Security is a focal point. Google described Agent Identity, a system that issues a cryptographic identifier to each agent so enterprises can track actions back to a discrete agent instance and maintain an audit trail of which agent performed which operation.

For risk-averse customers, Google built an Agent Simulation tool that lets teams run agents through staged, real-world scenarios before they go live. The simulation environment is intended to reveal failure modes, unsafe behaviors or data leaks without exposing production systems.

Operational controls extend into governance and DevOps. The platform includes orchestration primitives, role-based access and deployment workflows so organizations can version, test and roll out agent updates in a controlled way — addressing corporate concerns that agents may change behavior as models or prompts evolve.

No-code publishing and worker-facing apps

Once an agent is built and tested, it can be published to a Gemini Enterprise app for day-to-day use. That app is aimed at non-engineering staff: employees can run published agents or use low-code/no-code tools like Agent Studio and Agent Designer to compose or customize agents from templates.

In a demo, Google deployed multiple agents to coordinate an inventory and marketing workflow—reconciling stock, generating product descriptions and preparing campaign drafts. Google’s pitch: make agent orchestration feel like managing a small team, not writing bespoke automation scripts for every task.

Related Articles

Thomas Kurian said the new Agent Platform makes Gemini Enterprise an "end-to-end system for the agentic era," folding Vertex AI services together with security, testing and orchestration into a single enterprise workflow.