OpenAI Overhauled Its Agents SDK — Adds Sandbox Execution, Persistent Memory, and Bring-Your-Own Compute
OpenAI released a major update to its Agents SDK, adding native sandbox execution environments, configurable memory, MCP-style tool integrations, and support for eight compute providers — positioning it as a full agentic infrastructure stack, not just a wrapper around the chat API.
Original sourceOpenAI shipped what it's calling "the next evolution" of its Agents SDK, transforming it from a lightweight orchestration layer into a full-stack agentic development framework. The update adds sandbox-aware execution that lets agents run code safely in isolated environments, configurable memory for maintaining state across tasks, and a filesystem tool layer modeled on how Codex CLI operates.
The biggest structural change is native support for eight compute providers for sandboxed execution: Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel are all natively supported, alongside OpenAI's own infrastructure. This "bring your own sandbox" model gives enterprise developers full control over where agent code runs — a critical requirement for teams with data residency or compliance constraints.
The SDK also deepens MCP (Model Context Protocol) integration, with standardized tool connectors that work across the growing MCP ecosystem. This makes Agents SDK-based systems interoperable with the 97M+ install base of MCP tooling that's now standard across most major AI providers.
For developers, this represents a qualitative shift: the Agents SDK is no longer just a way to call GPT-4o in a loop. It's now a programmable agentic substrate with production-grade concerns like isolation, memory, and third-party compute baked in. The timing — alongside Codex 3.0's release — suggests OpenAI is consolidating its developer story around a coherent agentic platform, not just a model API.
The release also signals OpenAI's awareness that Google's ADK 2.0 and Anthropic's Claude Code ecosystem have been gaining ground among sophisticated agent builders. Whether the SDK's quality of life improvements close that gap will depend on how quickly the third-party integrations mature.
Panel Takes
The Builder
Developer Perspective
“Native E2B and Modal support out of the box saves weeks of integration work. The configurable memory model is what I've been manually implementing in every agent project. This update makes the SDK competitive with what ADK 2.0 shipped last week.”
The Skeptic
Reality Check
“Eight compute providers sounds impressive until you read the fine print — most require separate accounts and billing. 'Bring your own sandbox' means 'set up your own infrastructure.' This is still far from the one-click agentic deployment that the launch post implies.”
The Futurist
Big Picture
“Standardizing sandbox execution and memory interfaces at the SDK layer is exactly how you create a durable developer ecosystem. These aren't features — they're foundations. The AI agent landscape in 2027 will be built on these primitives being established right now.”