Google Cloud has launched an end-to-end solution for building and deploying autonomous AI agents at enterprise scale, and it has open-sourced every line of demo code shown at the event.
The Gemini Enterprise Agent Platform, introduced at Google Cloud Next 2026, is the company’s most structured effort to date to move AI agents out of experimental prototypes and into production environments.
The platform spans the entire lifecycle of agent development, from Agent Development Kits (ADKs) for building agents to serverless agent runtimes for deploying and scaling them without managing infrastructure.
Brad Calder, Google Cloud president of site reliability engineering, designed it to help developers create agents that “actively help users and complete tasks independently”.
Beyond development and runtime, the platform introduces administration infrastructure: a unique agent identity per agent instance, an agent gateway to enforce IAM policies, an agent registry for discovery, and an A2A protocol for agent-to-agent collaboration. Importantly, the platform supports models beyond Gemini Pro and Flash, including Anthropic’s Cloud via Google Cloud’s Model Garden.
To demonstrate the platform’s capabilities, Google’s developer Keynote used multi-agent simulation to plan a marathon in Las Vegas. The demo featured a planning agent, an evaluator and a simulator working in coordination, mapping and GIS tools, stateful memory, RAG integration, and dynamic interfaces built at runtime.
Security was woven directly into the architecture rather than implemented as an afterthought. The agent gateway acted as a proxy enforcing read-only policies on the Finance MCP server, preventing the planner agent from executing write operations or accessing the open Internet. The demo showed that governance is not optional in this stack; This is structural.
Viz co-founder Yinon Costica demonstrated how the platform integrates with cloud security tooling through a pair of specialized agents. A “Wiz Red Agent” identified an authentication bypass vulnerability running from an Internet-exposed entry point to sensitive data.
A “Wiz Green Agent” then recommended prioritizing fixes, upgrading IAM privileges, patching bypasses, and implementing AI guardrails. Cloud Code applied the changes, and Viz re-scanned to confirm the solution.
The sequence showed something important: AI agents are not only handling task execution but also proactive security improvements within the same platform.
Google made the entire demo codebase publicly available, including architectural guidance, labs, and developer credits. This move lowers the barrier for engineering teams that want to test multi-agent workflows without having to rebuild from scratch.
