Google Cloud executives and developer advocates used a marathon-planning simulation set in Las Vegas to demonstrate how Alphabet NASDAQ: GOOG is positioning its Gemini Enterprise Agent Platform for building “production-ready agents,” including tooling for development, runtime scaling, governance, observability, and security.
Google Cloud outlines the Gemini Enterprise Agent Platform
Brad Calder, President of Site Reliability Engineering for Google Cloud, opened the developer keynote by recapping the prior day’s introduction of the Gemini Enterprise Agent Platform, which he said is designed to help developers build autonomous agents that “proactively help users and complete tasks independently.”
Calder said the platform is powered by Gemini models, including Pro and Flash, and also supports other models “like Claude” through Google Cloud’s Model Garden. He described the platform’s components as spanning the full lifecycle of agent creation and operation, from the Agent Development Kit (ADK) to a serverless Agent Runtime for running and scaling agents. He also highlighted governance features such as “a unique Agent Identity” and policies enforced through an Agent Gateway, along with discovery and collaboration through an Agent Registry and an “A2A protocol.”
Calder added that Google is building its own applications on the platform to provide a “common shared context” across Gemini Enterprise, Workspace, and third-party marketplace agents, enabling agents to collaborate with shared context.
A marathon-planning demo to show multi-agent systems
Senior Director and Chief Evangelist Richard Seroter and Developer Relations Engineer Emma Twersky framed the keynote around a multi-agent simulation for planning a marathon route through Las Vegas, aiming to show how agent systems can simulate behavior rather than relying on hard-coded logic.
Seroter said the system used three primary agents:
- Planner agent: determines marathon routes
- Evaluator agent: assesses routes against business and community requirements
- Simulator agent: generates actors and randomized behaviors to model impact on the city
Twersky demonstrated a simulation app rendering the Las Vegas skyline and a proposed route at night, noting the app uses Angular for the base and 3D rendering and incorporates A2UI and “GenUI ideas” from Flutter. She showed features including a “Follow the leader” view of runners, placement of hydration stations and medical tents, simulated traffic, and reports that score prior simulation runs under “non-deterministic conditions.”
Seroter also said the full demo was available in a GitHub repository, describing it as an app attendees could reproduce themselves.
Building the planner: ADK, remote MCP tools, and skills
Mofi Rahman, Senior Developer Relations Engineer at Google Cloud, walked through building a planner agent from “agent idea to proof of concept” using ADK, Google Cloud remote MCP (model context protocol) servers, and Agent Runtime.
Rahman said the planner agent required instructions, mapping and geospatial skills, and tools for mapping and route calculation. He demonstrated using Agent Designer to create a basic agent with “low and no-code” workflows and preview behavior, then downloading Python code pre-populated with instructions as a starting point for ADK development.
To ground responses in real data, Rahman connected the agent to the Google Maps MCP server so it could retrieve information such as Las Vegas landmarks. He described skills as having YAML metadata and a Markdown body, allowing the agent to load specialized knowledge when needed. He highlighted three skills: mapping, GIS (including Python scripts to process GeoJSON), and a “race director” skill created by converting an internal Google Doc into a skill with Gemini. Rahman said the agent could be deployed to Agent Runtime in “4–5 minutes.”
Evaluation, A2UI, A2A/Registry networking, and context engineering
Casey West and Ivan Nardini showed how Google Cloud is trying to move from what West called “fragile, unpredictable agentic loops” to “a rigorously evaluated network of experts.” They described an evaluator sub-agent that uses a separate model and limited context to grade route plans, checking both non-deterministic criteria (such as community impact and prompt alignment) and deterministic criteria (including the requirement that a marathon be exactly 26 miles and 385 yards, or 42.195 km).
West also highlighted A2UI (agent-to-user interface), which he described as “an open standard created by Google,” enabling agents to generate dynamic UI components instead of returning only text. The pair then showed how A2A and Agent Registry connect agents without “brittle API code,” with West noting Google created A2A and “donated [it] to the Linux Foundation.” Nardini likened Agent Registry to “the DNS of your internet of agents,” automatically registering agents deployed to Agent Runtime.
Later, Lucia Subatin and Jack Wotherspoon focused on stateful behavior via sessions and memory. They said session and memory management could be added in “less than 20 lines of code,” attaching to an “enterprise-ready and fully managed memory service” called Agent Platform Memory Bank. They also demonstrated bringing in unstructured data via RAG by chunking documents with Document AI, processing with Lightning Engine for Apache Spark, and storing chunks in AlloyDB. They highlighted AlloyDB “auto-embeddings” and showed a semantic search example for rules on the Las Vegas Strip, including a restriction that “you can’t have a camel on public roads.”
Observability, autonomous operations, security governance, and Wiz integration
Megan O’Keefe demonstrated agent observability and Gemini Cloud Assist to debug a simulator agent issue. She said the simulator experienced high latency and crashes, and traced logs and tool calls through an agent runtime trace view. Using a Cloud Assist investigation, she said the simulator was failing Gemini model API calls due to a request error tied to exceeding the Gemini API’s “one million context token limit.” O’Keefe said Event Compaction in ADK—used to periodically summarize workflows—was not running frequently enough, and she demonstrated a fix by adding a token threshold parameter, then redeploying to Agent Runtime.
Bobby Allen then showed an infrastructure-focused workflow: converting a runners component from Cloud Run to Google Kubernetes Engine (GKE) and deploying a customized model based on “Gemma 4” in the same cluster, using Cloud Assist to generate infrastructure changes. He said a scaling issue emerged when running thousands of runners and that Cloud Assist identified model-loading bottlenecks, recommending Lustre over GCS FUSE for faster scale.
Security and governance were addressed through Agent Identity and Agent Gateway policies. Ankur Kotwal demonstrated using Agent Gateway as a proxy enforcing IAM policies, describing agent identities as “unique and immutable” per agent instance. He added a “Read Only Finance” policy to prevent the planner agent from using write-enabled tools on a finance MCP server, while keeping a separate policy blocking planner access to the open internet.
Wiz co-founder and VP of Product Yinon Costica presented Wiz’s “unified cloud security platform,” describing a security graph and agent-based workflows including a “Wiz Red Agent” for validating exploitability and a “Wiz Green Agent” for remediation. Costica said the Red Agent found an authentication bypass vulnerability along an attack path from internet exposure to sensitive data, and that the Green Agent suggested prioritized fixes including downgrading IAM privileges, patching the bypass, and enforcing AI guardrails. In a follow-on demo, they said Claude Code applied changes and Wiz rescanned the environment to confirm the issues were resolved.
In closing remarks, Twersky said Google had open sourced “every line of code” shown during the keynote, and Seroter said the repository includes architectural guidance, labs, and credits to help developers reproduce the project.
About Alphabet NASDAQ: GOOG
Alphabet Inc NASDAQ: GOOG is a multinational technology holding company headquartered in Mountain View, California. Formed in 2015 through a corporate restructuring of Google, Alphabet serves as the parent to Google LLC and a portfolio of businesses collectively known as "Other Bets." Google was originally founded in 1998 by Larry Page and Sergey Brin; Alphabet is led by CEO Sundar Pichai, who oversees Google and the broader company while the founders remain prominent shareholders and influential figures in the company's history.
Alphabet's core business centers on internet search and advertising, with Google Search and the company's ad platforms (including Google Ads and AdSense) generating the majority of revenue by connecting advertisers with consumers worldwide.
Featured Stories
This instant news alert was generated by narrative science technology and financial data from MarketBeat in order to provide readers with the fastest reporting and unbiased coverage. Please send any questions or comments about this story to contact@marketbeat.com.
Before you consider Alphabet, you'll want to hear this.
MarketBeat keeps track of Wall Street's top-rated and best performing research analysts and the stocks they recommend to their clients on a daily basis. MarketBeat has identified the five stocks that top analysts are quietly whispering to their clients to buy now before the broader market catches on... and Alphabet wasn't on the list.
While Alphabet currently has a Buy rating among analysts, top-rated analysts believe these five stocks are better buys.
View The Five Stocks Here
Click the link to see MarketBeat's list of seven stocks and why their long-term outlooks are very promising.
Get This Free Report
