Matt Brockman explains why running agent-generated code needs an isolated environment with controlled resources and accessAn agent permission boundary defines and enforces which resources and actions an automated agent is allowed to use.. The workshop uses snapshot-based sandboxesAn AI agent sandbox is an isolated execution environment that limits which files, processes, networks, credentials, and external systems an agent can access. and a capture-the-flag exercise to demonstrate runaway CPU, excessive memory use and disk growthPerformance profiling measures where a program spends time or computational resources so optimization can target observed bottlenecks., showing how to investigate rather than assume that isolation solves every operational problem.
Later exercises cover task runtimes, idle timeouts, orphan processes, persistent user-to-sandbox assignment and separate read-only template and writable runtime caches. Questions highlight tradeoffs in storage, networking, long-running services and GPU workloads. The examples are workshop demonstrations with stated limits, not a claim that every infrastructure use case is supported.
Watch on YouTube




