Building secure and scalable agents is hard. From multi-tenancy and noisy-neighbor risk to tool access and data boundaries, teams need more than a prompt and a container image, they need a deliberate execution model.

We will show what's available to run thousands of agents safely, focusing on open source, Kubernetes-based building blocks such as the CNCF Agent Sandbox, Agent Substrate, and Isola. We will cover their full lifecycle, provision, execute, observe, and tear down, while keeping workloads isolated. Sandboxes enforce resource limits, reduce blast radius, and give operators a consistent place to apply policy, though this adds trade-offs like state management and uneven resource use.

At Adobe we have optimized this so teams can run agents securely without needing to understand the sandboxing mechanism themselves.

In this session we will share the problems we hit running agents for real customers, how different projects address them, and what we would do differently next time.