OpenAI Agents SDK adds TypeScript support and sandbox agents

• OpenAI's dev account OpenAIDevs announcement said the Agents SDK is now available in TypeScript, and dkundel's v0.9.0 link post pointed developers straight to the JavaScript release that adds Sandbox Agents.
• The JavaScript release turns sandboxing into a first-class SDK surface: the v0.9.0 release, dkundel's post, and OpenAIDevs announcement all frame it around persistent workspaces, command execution, repo editing, and resume support.
• OpenAI's earlier launch post, linked by algo_diver's repost and TheRealAdamG's link post, describes the bigger shift as a model-native harness with sandbox-aware orchestration, configurable memory, Codex-like filesystem tools, MCP, and skills.
• The new TypeScript docs show SandboxAgent using the same run() flow as the rest of the SDK, while the runtime docs add knobs for sandbox client, live session, snapshot, manifest override, and concurrency in a single TypeScript docs surface.
• Sam Altman's short sama reaction post makes the subtext plain: OpenAI wants this update read as more than a language port, even if the immediate shipping news is the JavaScript catch-up that cedric_chee's repost helped amplify.

You can read OpenAI's launch post, skim the openai-agents-js v0.9.0 release, and browse the new TypeScript docs. The interesting bits are lower level than the headline: sandbox runs now distinguish sessionState from snapshots in the sandbox client docs, manifests can materialize local files or Git repos, and the runner config exposes sandbox client and concurrency controls in the run config guide.

The public news is simple: OpenAI finally shipped the upgraded Agents SDK surface to TypeScript, not just Python. The TypeScript homepage shows SandboxAgent, Manifest, and a local sandbox client wired into the standard run() flow, so the JavaScript version is not a side package. It is the same core abstraction with a sandbox-specific agent class.

The v0.9.0 release fills in the practical pieces that matter for JS teams. According to the release notes, Sandbox Agents add:

• persistent workspaces
• capabilities for files, shell, patches, images, memory, and compaction
• serialized run state for resuming work
• manifests for synthetic files, local mounts, and Git repos
• hosted and local sandbox backends

That is a broader port than "TypeScript support" implies. It moves the JS SDK from orchestration glue toward a coding-agent runtime.

The release notes and docs are most concrete around workspace state. The sandbox client guide splits persistence into two different mechanisms:

• sessionState reconnects to the same backend sandbox session.
• snapshots start a fresh session from saved workspace contents.
• RunState can carry runner-managed sandbox state when a larger workflow pauses or resumes.

That split matters because OpenAI is not only offering an ephemeral container. It is exposing a longer-lived workspace model that can survive across runs.

The provider list is also unusually wide for a first JavaScript cut. The v0.9.0 release names Unix-local and Docker clients, plus hosted integrations for Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel. dkundel's post links directly to that inventory, while TheTuringPost's overview separately called out that sandbox agents get access to files, repos, and terminal commands.

OpenAI's own writeup spends less time on language support than on the harness idea. In the launch post, the company describes a model-native harness plus native sandbox execution, then adds configurable memory, sandbox-aware orchestration, Codex-like filesystem tools, MCP integrations, and skills on top.

That framing shows up again in the JavaScript runtime docs. The run config guide says a sandbox run can be configured with a client, live session, serialized session state, snapshot, manifest override, and concurrency limits. OpenAI is packaging the loop, the workspace, and the durable execution controls as one SDK surface.

Sam Altman's entire public take was four words: "Agents SDK 2.0 is underrated." The better reading, backed by TheRealAdamG's diagram post and TheRealAdamG's link post, is that OpenAI thinks the harness layer is the strategic part. TypeScript support is the visible ship. The attempt to standardize sandboxed, long-horizon agent execution is the bigger bet.