Google Goes Agentic

Google is turning search and office software into AI agents that can plan, gather sources, and return finished reports with less back-and-forth from users. (blog.google) On April 21, Google introduced Deep Research and Deep Research Max, two Gemini tools built for “long-horizon” research tasks across the web and private data sources. Google said the new versions are powered by Gemini 3.1 Pro and add Model Context Protocol support plus native charts and visualizations. (blog.google, blog.google) These systems are designed to break a big assignment into smaller steps, search repeatedly, compare findings, and assemble results into a report. Google said the agents can work across web sources or custom sources instead of answering with a single pass like a standard chatbot prompt. (blog.google) Google is also moving the same pattern into Workspace. In March, the company said Ask Gemini in Drive would let users get detailed answers grounded in files stored in Drive, along with Gmail, Calendar, and Chat context. (workspace.google.com, blog.google) That shift puts Google’s core products closer to a model where AI does the legwork inside the company’s own ecosystem. Search already uses AI Overviews for quick summaries, while Drive and Docs are being updated to pull from workplace documents and draft fuller outputs. (blog.google, workspace.google.com) Google tied the software push to new hardware on April 22 at Cloud Next 2026. The company announced eighth-generation Tensor Processing Units, with TPU 8t for training and TPU 8i for inference and reinforcement learning, and said the split reflects the different demands of agent-style systems. (cloud.google.com, cloud.google.com) Inference is the stage where a trained model produces answers for users, and agent systems do more of it because they loop through searches, tool calls, and checks before replying. Google said real-time inference needs high memory bandwidth and low latency, which is why TPU 8i is being positioned for large-scale serving. (cloud.google.com, cloud.google.com) At Cloud Next, Google framed the broader strategy as the “Agentic Enterprise,” with Gemini Enterprise combining models, user interfaces, development tools, and system access for multi-step work. The company’s pitch is that businesses will delegate outcomes to agents rather than use chatbots for one question at a time. (cloud.google.com, cloud.google.com) The immediate test is whether users trust these agents to do research without hiding errors behind polished summaries. Google’s answer, for now, is to pair bigger reasoning models with more source access, tighter links to company data, and hardware built for repeated AI decision loops. (blog.google, cloud.google.com)