Build Fast, Stay Safe: Sandbox Architecture for AI Workloads

You have an idea for an AI agent, a new RAG pipeline, or some other experiment, but getting a safe environment to build it takes longer than the experiment itself. Between account provisioning, IAM policy reviews, and the fear of runaway costs, most experiments never leave the whiteboard. In this session, you’ll learn how teams set up isolated, disposable experiment environments with guardrails built in from day one, using production-ready, fully supported open-source patterns on AWS. We’ll walk through the architecture behind self-service sandboxes: how service control policies, identity boundaries, and budget automation work together to give builders freedom without risk. Then we’ll apply these patterns to AI workloads, where foundation model costs and compute sprawl make governance even more critical. Whether you’re a developer who just wants a safe place to break things, or an architect designing your team’s experimentation workflow, you’ll leave with practical patterns you can take back and implement. 🧠 What you’ll learn Why traditional account provisioning slows innovation, and how self-service sandboxes fix it How to use Service Control Policies (SCPs) to set safe boundaries for every experiment Identity and permission boundaries that let builders move fast without broad access Budget automation and cleanup so experiment environments never turn into surprise bills Governance patterns for AI workloads: controlling foundation model usage, GPU and compute sprawl, and cost visibility An end-to-end reference architecture using open-source, production-ready patterns on AWS 🎤 Speakers Luis Guirigay – Worldwide Head of Cloud Infrastructure Solutions, AWS Luis leads AWS’s global cloud infrastructure solutions and works with organizations on how to build and govern at scale. Peter DeVries – Product Management, Agentic AI Adoption, AWS Peter works on accelerating the adoption of agentic AI, helping teams move AI ideas from prototype…