OpenAI updated its Agents SDK with sandboxing, in-distribution harnesses, and workspace isolation to help enterprises deploy safer, more controlled AI agents.
OpenAI released a major update to its Agents SDK, adding sandboxed computer environments that restrict agents to specific files and tools within a workspace. The update also introduces an in-distribution harness for frontier models, enabling both deployment and testing of advanced agents. Initially launching in Python only, with TypeScript support planned later. Available now to all API customers at standard pricing.
The new SDK gives developers a legit isolation layer — agents can now be scoped to specific files, tools, and workspace contexts without touching the broader system. The in-distribution harness means you can test frontier model agents in the same environment they'll run in production. Python-first, TypeScript coming later — if you're building in TS, you're waiting.
Audit your current agent setup this week: if your agent has unrestricted filesystem or tool access, swap in the new sandbox integration via the OpenAI Agents SDK and measure whether scoped permissions break any existing flows before your next deployment.
Suggested exercise · not independently tested. Access requirements and costs may vary.
Install or upgrade the SDK: pip install --upgrade openai-agents
Paste this into a new Python file: from agents import Agent, Runner, Sandbox; agent = Agent(name='sandbox-test', instructions='List files in /workspace only', sandbox=Sandbox(allowed_paths=['/workspace'])); result = Runner.run_sync(agent, 'What files can you see?'); print(result.final_output)
Run it and observe whether the agent respects path restrictions — you'll see either a scoped file list or an access-denied response confirming isolation is active
Tags
Related