Google AI Updates: September 23, 2026

1. Google Open Sourced AX, an Agent Orchestrator That Treats Agents as Stateful Actors

Google. AX is an Apache 2.0 orchestrator and declarative runtime, hosted at agentexecutor.io and on GitHub as google/ax, built on the Agent Substrate execution runtime. The design premise is that agents are neither stateless microservices nor batch jobs: they burn compute during reasoning and tool execution, then idle for long stretches waiting on model responses or human input. AX runs each session as an isolated actor sandbox with CPU and memory bounds, checkpoints and suspends it when idle, and claims sub-second resume with no cold start, multiplexing dozens of tasks onto shared host workers. Four declarative primitives: Task, Workspace, Gateway, and Model. Source

2. Colab Premium Compute Is Now Bundled Into Google AI Subscriptions

Google. Google AI subscribers get priority access to faster accelerators and more powerful Colab machines, and Google AI Ultra subscribers additionally unlock Premium GPU access and uninterrupted background execution, so long training runs no longer require keeping a browser tab open. Existing standalone Colab subscriptions are unchanged and the benefits stack. The bundle also covers expanded cloud storage and premium access to Google Antigravity and AI Studio. Source

3. BigQuery Data Agents Can Now Be Published Into Gemini Enterprise

Google. A BigQuery data agent can be registered with Agent Registry and imported into Gemini Enterprise using default Google-managed credentials, provided BigQuery and Gemini Enterprise sit in the same Google Cloud project with a matching Agent Gateway configuration. That removes the service account plumbing that previously stood between a data agent and an enterprise-wide agent surface. Source