OpenAI Adds Sandboxing to Agents SDK for Controlled Agent Workspaces
OpenAI updated its Agents SDK this week with sandboxing support, giving agents isolated workspaces to operate in. The change targets enterprise use cases where containment matters more than flexibility.
What Changed
The new sandbox integration allows agents to work in siloed environments. File and code access is scoped to specific operations only. Nothing bleeds across task boundaries.
Alongside the sandbox, OpenAI added an in-distribution harness for frontier models. Agents running in this harness get access to files and approved tools within their workspace, nothing outside it.
Both capabilities are designed for long-horizon, multi-step tasks. The claim is that tighter scoping makes agents safer to deploy on extended workflows.
Python First, TypeScript Later
The new harness and sandbox features launch in Python only. TypeScript support is planned but has no announced timeline. OpenAI says it is also working to bring code mode and subagent capabilities to both languages.
The Python-first pattern is consistent with how OpenAI has shipped recent SDK updates. TypeScript tends to follow a few weeks behind.
Pricing and Access
The new features are available to all customers through the standard API. No separate tier, no waitlist. Standard pricing applies.
The Practical Read
Sandboxing for agents is not a new idea. The interesting part here is the in-distribution harness, which constrains what frontier models can reach during agentic runs. Whether this meaningfully reduces real-world failure modes depends on how the access controls are implemented, which OpenAI has not detailed publicly.
For teams already using the Agents SDK, this is worth testing. For teams evaluating agent frameworks, it moves the SDK closer to what production deployments actually need.
Source: Techcrunch