4.1.4. AI and ML Development Environments
💡 First Principle: Google Cloud's AI/ML development surface spans from raw accelerator hardware up through fully managed development environments, and operational tasks here mostly involve attaching the right resource to the right workload and choosing the right environment for the development stage you're in.
Attaching GPUs and TPUs to a workload (a Compute Engine VM, a GKE node pool, or a training job) is an explicit operational configuration step, not automatic. Deploying an agent to Agent Runtime on the Gemini Enterprise Agent Platform hosts a built AI agent in a managed, scalable runtime rather than self-hosting it on general-purpose compute. Managing notebooks in Gemini Enterprise Agent Platform Workbench (the managed Jupyter-notebook environment for ML development) and in BigQuery (notebooks embedded directly against BigQuery data for in-place analysis) supports the iterative, exploratory phase of ML and data work. Cloud Workstations provides fully managed, browser-accessible development environments for general software engineering — pre-configured, consistent, and centrally administered, reducing "works on my machine" drift across a development team.
⚠️ Exam Trap: A scenario asking for a fully managed, standardized development environment for a general software engineering team (not specifically ML-focused) is pointing at Cloud Workstations, not Workbench notebooks — Workbench is scoped to ML/data-science notebook workflows, while Cloud Workstations is the general-purpose managed dev environment.
Reflection Question: A data science team wants to iteratively explore a BigQuery dataset using a notebook interface without exporting the data elsewhere first. Which feature supports that directly?
