OpenAI · 2026-04-15 · notable
OpenAI Agents SDK v0.14.0 — Sandbox Agents with Persistent Isolated Workspaces
OpenAI Agents SDK v0.14.0 adds Sandbox Agents in beta — persistent, isolated workspaces with filesystem access, shell execution, and cloud storage mounts. Supports local, Docker, and hosted providers (E2B, Modal, Vercel, Runloop).
OpenAI's Agents SDK now gives agents a persistent, isolated workspace — filesystem, shell, cloud storage, all in one run.
Key specs
| Version | 0.14.0 |
|---|---|
| GitHub stars | 20.8k |
| Sandbox backends | 7+ (E2B, Modal, Vercel, Runloop, Cloudflare, Docker, local) |
What is it?
Sandbox Agents is a new beta feature in the OpenAI Agents SDK (v0.14.0) that gives an agent a managed workspace it can use across multiple runs. Instead of rebuilding state from scratch on every invocation, the agent gets a real environment with files, shell access, and cloud storage it can read from and write to directly. Released as open-source Python under the MIT license.
How does it work?
A SandboxAgent wraps a standard Agent with a manifest (declaring which files, Git repos, environment variables, and remote storage buckets to stage) and a SandboxRunConfig (specifying the execution backend and session lifetime). Supported backends include local Unix, Docker containers, and hosted providers — E2B, Modal, Cloudflare, Vercel, Runloop, Daytona, and Blaxel. The SDK handles staging, snapshotting, and session resumption automatically, so agents can pick up mid-task where they left off. Cloud storage mounts support S3, GCS, Azure Blob Storage, and Cloudflare R2.
Why does it matter?
Long-horizon coding agents — write, test, commit, iterate — have relied on ad-hoc file staging and manual sandbox wiring. Sandbox Agents standardizes this into the SDK itself with a declarative manifest and swappable backends, so teams can use their own infrastructure without building the plumbing from scratch. Because the harness is open-source, it can be audited, extended, or self-hosted.
Who is it for?
Python developers building long-horizon or multi-step agents
Try it
pip install "openai-agents>=0.14.0"