OpenAI Updates Agents SDK to Support Secure Sandboxes and Durable Execution

OpenAI

Apr 15, 2026 · Updated Apr 25, 2026

OpenAI updated its Agents SDK with a model-native harness and native sandbox execution to help developers build production-grade autonomous agents. By separating the agent's logic from its execution environment, the system enables durable runs that can recover from crashes and execute code safely in isolated workspaces.

OpenAI released a major update to its Agents SDK, introducing a model-native harness (the core engine managing the agent loop) and native sandbox execution. The harness is optimized for frontier models to handle long-horizon tasks, while the sandbox provides a secure environment where agents can safely run commands, edit files, and install dependencies.

Building reliable agents currently requires complex infrastructure to manage state and security. This update standardizes the agent stack by decoupling the harness from compute, which prevents credential leaks and enables durable execution. If a sandbox fails, the harness can rehydrate (restore state from a snapshot) in a fresh container and continue without losing progress.

You can now build agents that integrate with MCP and use AGENTS.md for instructions. The SDK supports sandbox providers like E2B and Vercel out of the box. These features are generally available to API customers in Python, with TypeScript support coming soon.

View the full update on openai.com

OpenAI Developers

@OpenAIDevsApr 15

Build long-running agents with more control over agent execution. New capabilities in the Agents SDK: • Run agents in controlled sandboxes • Inspect and customize the open-source harness • Control when memories are created and where they’re stored https://t.co/zPyuLup6b6

1151.2k

View on X

Every HeadsUpAI update is written based on its original source and reviewed before it's published. Read our editorial standards →

See all AI news & updates from OpenAI →

Keep reading

OpenAI Secures Agent Workflows by Separating Credentials from Execution Environments

OpenAI updated its Agents SDK to support isolated execution environments through a native Vercel Sandbox integration. This architecture keeps sensitive credentials in the orchestration layer while allowing agents to run code in secure MicroVMs. It enables developers to build autonomous systems that can safely execute untrusted code without risking host system exposure.

Cloudflare Launches Sandboxes to Give AI Agents a Persistent Computer

CloudflareApr 13

Cloudflare Launches Sandboxes to Give AI Agents a Persistent Computer

Cloudflare Sandboxes are now generally available, providing AI agents with isolated, persistent environments to execute code and manage development servers. This infrastructure shift allows agents to perform multi-step engineering tasks like human developers while maintaining security through programmable credential injection.

OpenRouter Agent SDK Adds Human-in-the-Loop Tools to Pause High-Stakes AI Tasks

OpenRouterMay 8

OpenRouter Agent SDK Adds Human-in-the-Loop Tools to Pause High-Stakes AI Tasks

OpenRouter introduced human-in-the-loop controls to its Agent SDK, allowing developers to build agents that pause for review on sensitive actions. The update includes hooks to auto-resolve routine tasks while surfacing high-stakes decisions to a human interface. This standardizes how autonomous agents ask for help without losing conversation state.

NVIDIA Hardens OpenShell With Agent-Driven Policies and Sandbox Resource Limits

NVIDIAMay 15

NVIDIA Hardens OpenShell With Agent-Driven Policies and Sandbox Resource Limits

NVIDIA released OpenShell v0.0.41, adding agent-driven policy management and granular resource controls for its secure AI agent runtime. The update introduces workspace-boundary checks and custom certificate support to prevent sandbox escapes in enterprise environments.