Google Cloud unveils 13 Gemini Enterprise agent demos

Google Cloud has published 13 demonstrations for its Gemini Enterprise Agent Platform, showing developers how to build, deploy, govern and assess AI agents on the platform.

The demonstrations centre on Google's code-first Agent Development Kit, or ADK, and span workflows from basic conversational agents to multi-agent systems operating across programming languages and frameworks. The material is intended to show architectural patterns and implementation approaches developers can apply in production.

The release adds practical examples to a product Google introduced earlier this year for building and managing enterprise AI agents. It also highlights the company's Agents CLI tool, which can be installed in existing coding assistants and used to scaffold, evaluate, deploy and monitor agents from within a developer's editor.

Build stage

Several demonstrations focus on early-stage development. One introduces a first agent built with ADK, including environment setup, configuration and testing through both a command-line interface and a web interface.

Another walks through an event-driven expense approval agent that uses a graph-based workflow API. Lower-value expenses are approved automatically, while higher-value claims go through security screening, Gemini-based compliance analysis and human review before completion.

Google also highlights examples of connecting agents to data through the Model Context Protocol, or MCP. One demonstration shows how to build tools for querying BigQuery, searching files and calling application programming interfaces, with the aim of making them reusable across suppliers and software frameworks.

A separate example covers Agent-to-UI, or A2UI, in which an agent generates interface components such as layouts, charts and menus in real time as a conversation progresses.

Operational use

The second group of demonstrations turns to deployment and scale. One tutorial covers a stateful data science agent linked to BigQuery that stores user preferences across sessions through what Google calls Memory Bank before being deployed to Agent Runtime, the service previously known as Agent Engine.

Another focuses on long-running agents designed for workflows that can last days or weeks. The example uses durable state machines, event-driven idle handling and checkpoint-and-resume patterns with persistent sessions so tasks can continue after interruptions or restarts.

There is also a deployment example for the expense agent using Agents CLI. This setup adds observability through Cloud Trace and Cloud Logging, routes analytics to BigQuery Agent Analytics and registers the agent in Agent Registry after deployment.

The final scaling example adds a front end for managers through a dashboard on Cloud Run. The design connects to Agent Runtime through an authenticated Pub/Sub pipeline and lets managers resume paused review sessions from a browser.

Security controls

Governance and security form a separate section of the release. One demonstration presents a test-first development process for a shopping assistant and includes a STRIDE threat model, a Semgrep pre-commit hook and a control designed to block risky actions before execution.

Another addresses runtime access control through Agent Gateway. In that model, agents use unique identities and make outbound calls through authentication and authorisation checks, while Model Armour inspects content for prompt injection attempts and possible data leakage.

These examples reflect a broader shift in enterprise AI from isolated prototypes to systems that require lifecycle management, access rules and auditability. Large organisations have increasingly sought ways to govern agents that interact with corporate data, external tools and human decision-makers.

Evaluation tools

The final set of demonstrations focuses on optimisation and interoperability. One tutorial introduces a five-stage evaluation cycle covering data preparation, inference, grading, failure analysis and targeted optimisation. The grading stage uses adaptive AutoRaters based on methods Google uses internally for model and agent evaluation in partnership with DeepMind.

Another example shows a cross-language compliance pipeline in which a Python-based agent extracts contract terms with Gemini and a Go-based agent validates them against internal policy. The services are connected through the Agent-to-Agent, or A2A, protocol and orchestrated through ADK.

The last demonstration expands that idea to multiple frameworks, with an ADK control layer delegating planning to LangGraph and execution to CrewAI through A2A. The design allows the system to re-plan automatically if a step fails.

The collection offers a clearer picture of how Google wants enterprises to use Gemini beyond chat-style assistants, particularly in workflows involving approvals, compliance checks, data access and software orchestration. It also shows the company promoting open protocols such as MCP and A2A to connect systems built in different languages and on different stacks.

Google says developers can install Agents CLI in their preferred coding agent and use plain-language prompts to start building on the platform.