Google's full-stack agent push

Google used its Cloud Next event on April 22 to argue that AI agents are becoming a business product, not just a model demo. Google’s pitch was a bundled stack that runs from office software to cloud chips. (reuters.com) (blog.google) An AI agent is software that can take a goal, pull data from tools, and complete a series of steps with less human prompting than a chatbot. At Cloud Next in Las Vegas, Google packaged that idea into what it calls the Gemini Enterprise Agent Platform and added Workspace Studio so companies can build agents inside Gmail, Docs, Sheets, Drive, Meet and other Google apps. (thenextweb.com) (blog.google) Google also pushed a system for agent-to-agent messaging, called Agent2Agent or A2A, which is meant to let one company’s software agent hand work to another across different apps and vendors. Google said the protocol now has backing from more than 150 organizations, up from more than 50 partners when it was introduced on April 9, 2025. (developers.googleblog.com) (cloud.google.com) The company said A2A is not meant to replace Anthropic’s Model Context Protocol, or MCP, which is used to connect models to tools and data sources. Google’s framing is that MCP helps an agent reach tools, while A2A helps separate agents coordinate work across systems and organizations. (thenextweb.com) (cloud.google.com) That distinction fits Google’s larger sales pitch to enterprise customers, who usually run software from many vendors and need new systems to fit old ones. Reuters reported that Google is trying to turn agents into a money-making cloud product as it competes with Microsoft, Amazon, OpenAI and Anthropic for corporate AI spending. (reuters.com) (blog.google) Google tied that software story to hardware. The company announced two eighth-generation Tensor Processing Units, splitting the line into TPU 8t for training models and TPU 8i for inference, the step where a trained model produces answers for users. (techcrunch.com) (blog.google) Google said TPU 8i is tuned for the fast response times that agents need when they carry out multi-step tasks, while TPU 8t is built for larger-scale model training. TechCrunch reported that Google is still offering Nvidia chips in its cloud, even as it uses custom silicon to lower costs and control more of the AI stack. (blog.google) (techcrunch.com) Google paired the announcements with adoption numbers meant to show that the market is moving from experiments to production. The company said nearly 75% of Google Cloud customers now use its AI products, 330 customers processed more than 1 trillion tokens in the past 12 months, and its first-party models now handle more than 16 billion tokens per minute through direct application programming interface use. (blog.google 1) (blog.google 2) The open question is whether companies want one vendor to supply the model, the agent framework, the office software and the chips underneath. Google’s answer on April 22 was that the easiest way to run agents at scale is to buy the whole system from one place. (reuters.com) (thenextweb.com)