Skip to main content
Back

OpenAI's Agents SDK Gives AI Workers a Real Persistent Workspace

OpenAI's Agents SDK gains Sandbox Agents in v0.14 -- giving AI agents a real filesystem, shell access, and resumable workspace to tackle long-horizon tasks

Repo, Full Access
  • OpenAI released Agents SDK v0.14 with Sandbox Agents -- giving AI agents a real persistent filesystem, shell, and resumable workspace.
  • Agents now run in isolated environments (local, Docker, or 7 hosted providers: Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, Vercel) without custom infrastructure.
  • A Codex-style apply_patch tool lets agents edit files via diffs instead of full rewrites, cutting token use and hallucination risk.
  • The SDK supports 100+ LLMs via Chat Completions API -- not just OpenAI models -- enabling mixed-model pipelines.
  • Sandbox Agents are in beta (Python only for now); TypeScript support is planned but not yet available.
  • OpenAI's Assistants API is being deprecated mid-2026; teams should migrate to the Responses API now.

Most agent frameworks are good at talking about work. The updated OpenAI Agents SDK is now built to actually do it. The biggest addition in version 0.14 is Sandbox Agents -- a new primitive that gives an agent a real, persistent workspace where it can read and write files, run shell commands, apply code patches, and pick up where it left off if the task gets interrupted. This is the SDK's largest overhaul since launch, and it closes the gap between agent prototypes and production deployments.

The gap this fills

The gap between a flashy agent demo and an agent that can actually run in production has been almost entirely about infrastructure -- not model capability. Before this update, teams building long-running agents had to wire together their own execution environments, manage file staging, handle sandbox lifecycle, and deal with state loss when containers restarted. Developers need more than the best models to build useful agents -- they need systems that support how agents inspect files, run commands, write code, and keep working across many steps.

The release of the OpenAI Agents SDK 0.14 marks a definitive shift in how developers build and deploy autonomous systems, moving beyond the simple prompt-and-tool loops of 2025 with a robust infrastructure layer designed for long-horizon tasks and secure code execution. The SDK itself started as a production-ready successor to Swarm, OpenAI's earlier experimental multi-agent framework, and has been steadily adding structure ever since.

What Sandbox Agents actually are

Modern agents work best when they can operate on real files in a filesystem. Sandbox Agents in the Agents SDK give the model a persistent workspace where it can search large document sets, edit files, run commands, generate artifacts, and pick work back up from saved sandbox state. You keep the normal Agent and Runner flow and add three new pieces: a Manifest describing the workspace, a set of Capabilities for sandbox-native tools, and a