OpenAI Cookbook Shows Sandboxed Code Migration Agents

OpenAI published a cookbook example demonstrating how to use its Agents SDK to run code migrations inside isolated sandboxes, each handling a shard of the codebase independently rather than producing one massive PR. The architecture keeps orchestration and credentials in a trusted host process while sandboxes only see task-scoped files and shell commands. Agents follow a MIGRATION.md brief, validate baselines before editing, run tests after patching, and return typed results with audit logs. The sandbox layer swaps between Docker, E2B, and Cloudflare without changing agent code. Practical migration patterns like this have been conspicuously missing from most agent SDK launches - Anthropic, LangChain, and others have focused more on general-purpose orchestration than showing opinionated, end-to-end workflows.