Overview
- Google introduced the Gemini Enterprise Agent Platform at its Next conference, folding Vertex AI into a new stack built to run fleets of stateful agents that can plan tasks, call tools, and keep memory over days.
- The platform centers on four pillars—Build, Scale, Govern, Optimize—with tools like Agent Studio and an Agent Development Kit for design and code, a long‑running Agent Runtime with a Memory Bank, and built‑in evaluation and tracing for quality checks.
- Governance is a headline feature, with cryptographic Agent Identity for audit trails, an Agent Registry as a single source of truth for approved agents and tools, and an Agent Gateway that enforces policies and blocks prompt‑injection and data leaks; Google is also backing standards such as the Model Context Protocol for safe tool access and an Agent‑to‑Agent protocol for inter‑agent handoffs.
- Google detailed an Agentic Data Cloud that auto‑indexes new documents through a Knowledge Catalog, adds graph‑aware retrieval using BigQuery Graph so agents can follow relationships, and lets teams query data across AWS and Azure via a Cross‑Cloud Lakehouse based on Apache Iceberg and federation that is in preview.
- For scale and cost, Google split its new TPU v8 into training‑focused 8t and inference‑focused 8i and said capacity rolls out later in 2026, paired with Nvidia NVL72 options, while early customers such as Comcast, Color Health, L’Oréal, Payhawk, and PayPal test agent workflows as no‑code tools like Workspace Studio and Wiz’s red‑team and fixer agents arrive with several features still in preview.