OpenAI released a major update to its Agents SDK on April 15, 2026, adding native sandbox execution, a model-native harness for file and tool operations, and durable execution with automatic snapshotting. The update targets the production gap that has kept many enterprise agent projects stuck at the prototype stage.
Sandbox Execution Across Seven Providers
The SDK now supports sandbox execution out of the box, letting agents run in controlled environments with defined file access, tool permissions, and dependencies. Developers can use built-in integrations with Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel, or bring their own sandbox provider.
A new Manifest abstraction lets developers describe the agent’s workspace declaratively: mounting local files, defining output directories, and pulling data from AWS S3, Google Cloud Storage, Azure Blob Storage, and Cloudflare R2. The workspace definition stays consistent from local development to production deployment, according to OpenAI’s announcement.
“This launch, at its core, is about taking our existing Agents SDK and making it so it’s compatible with all of these sandbox providers,” Karan Sharma from OpenAI’s product team told TechCrunch.
Model-Native Harness
The harness component aligns agent execution with how OpenAI’s frontier models perform best. It includes configurable memory, Codex-style filesystem tools (shell execution, apply-patch file edits), MCP-based tool use, progressive disclosure via skills, and custom instructions through AGENTS.md files.
OpenAI framed this as solving a real tradeoff in the current tooling landscape: model-agnostic frameworks are flexible but underutilize frontier model capabilities, while managed agent APIs simplify deployment but constrain where agents run and how they access sensitive data. The harness is designed to sit between those extremes, offering turnkey infrastructure that developers can adapt to their own stack.
Durable Execution and Security Architecture
The SDK externalizes agent state so that losing a sandbox container does not lose the run. Built-in snapshotting and rehydration allow the SDK to restore agent state in a fresh container and continue from the last checkpoint if the original environment fails.
OpenAI explicitly noted the security architecture assumes prompt injection and exfiltration attempts will occur. Separating the harness from the compute environment keeps credentials out of spaces where model-generated code executes.
Availability
The new capabilities are generally available to all API customers at standard pricing based on tokens and tool use. The harness and sandbox features launch first in Python, with TypeScript support planned for a future release. OpenAI said it is also working on bringing code mode and subagent capabilities to both languages.
The timing places OpenAI’s SDK update alongside a broader industry push toward enterprise agent infrastructure, with Cloudflare, Microsoft, and AWS all shipping agent execution and governance tools in the same week.