OpenAI added native sandbox execution and a model-native harness to its Agents SDK, enabling secure, persistent agents that operate across files and tools.
OpenAI released an update to its Agents SDK introducing native sandbox execution, which isolates agent code runs in a secure environment without additional infrastructure setup. A new model-native harness allows agents to persist across longer task horizons, managing files and tools more coherently. The update targets developers building production-grade agentic workflows that require reliability and security at scale. No pricing changes were announced, and the SDK update is available now via the OpenAI Python package.
The native sandbox removes the single biggest friction in production agentic systems: you no longer need to wire up your own containerized execution environment or worry about agent code escaping its scope. The model-native harness means session state and tool context persist across turns without you building your own memory layer. This directly reduces the boilerplate required to build file-handling, multi-tool agents by a meaningful margin.
Pull the latest Agents SDK this week and replace your existing subprocess-isolated agent runner with the native sandbox — benchmark whether your p95 task completion latency drops and whether error containment improves on a file-manipulation workflow.
Run: pip install --upgrade openai-agents in your terminal
Tags
Also today
Signals by role
Also today
Tools mentioned