OpenAI Agents SDK tutorial: Docker to Modal sandboxes

Steve Coffey demonstrates a conference task tracker built with the OpenAI Agents SDK. He describes the SDK as an open source framework with a Codex-style harness that separates the agent loop from its execution sandbox. The session covers Python and TypeScript support, shell commands, file editing and context compaction.

The demo starts with Docker on a laptop using a Python 3.12 image. A program editor agent receives uploaded documents, inspects them and edits their contents. Coffey then loads a conference editing skill from a Git repository. Filesystem snapshots let the task resume in a new container, while saved run state retains the conversation. Local sandbox execution here accompanies API calls; the demo does not establish local model inference or offline operation.

Next, Coffey moves execution to Modal and stores snapshots in Cloudflare R2. Custom function tools update task assignments and status, with a human approval step before marking work done. He also demonstrates passing work between agents and mounting an R2 bucket through an S3-compatible interface.

The Q&A explains SDK-managed versus user-managed sandbox lifetimes. The speakers also discuss a storage tradeoff: copying files adds startup work but speeds subsequent reads, while mounting avoids that initial copy and introduces network access during execution.