OpenAI updated the Agents SDK on April 15 with a set of capabilities that together are the clearest ratification yet of the standards the agent ecosystem has converged on over the past year. The update ships a model-native harness with configurable memory, sandbox-aware orchestration, and "Codex-like" filesystem tools. It standardizes on five primitives developers have been wiring up by hand: tool use via MCP, progressive disclosure via skills, custom instructions via AGENTS.md, shell-based code execution, and apply-patch file edits. Sandbox execution is supported natively with bring-your-own integrations for Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel. Python ships first, TypeScript follows. The interesting part is not any single feature but the shape of the aggregate move.
Five standards just got ratified in one release. AGENTS.md, a markdown file at project root that supplies custom instructions to the agent, is now an explicit OpenAI primitive alongside Anthropic's CLAUDE.md and Google Android's SKILL.md pattern from yesterday. The naming convention has standardized on "agent-legible file at project root," and with OpenAI adopting AGENTS.md the convention is effectively settled. MCP, which Anthropic shipped in 2024 and which a large third-party ecosystem adopted, is now a first-class OpenAI citizen. That was the last major lab holdout; MCP is the de-facto standard for tool interfaces. Skills, which Google used for the Android SKILL.md work, are now a named OpenAI SDK primitive for "progressive disclosure" of agent instructions. Shell-based code execution and apply-patch file edits are the two execution primitives that emerged from Claude Code and Cursor, and OpenAI is adopting them without renaming. Sandbox providers plural signals something different. OpenAI is not trying to own the sandbox layer; it is standardizing the interface and letting a marketplace develop. Cloudflare's Code Mode, which I covered yesterday, fits directly into this model.
The shape of this update is OpenAI moving from "we invent the standards" to "we adopt what the ecosystem has already chosen." That is a meaningful stance change. For the last two years the default assumption was that each lab would ship its own proprietary agent primitives and compete on the quality of those primitives. What has actually happened is convergence: MCP for tools, CLAUDE.md and AGENTS.md for instructions, SKILL.md and skills for progressive disclosure, sandbox abstractions for code execution. The labs that try to ship proprietary versions of these find that the third-party ecosystem routes around them. OpenAI ratifying the ecosystem standards is the pragmatic move. It signals that the agent-SDK layer is no longer a meaningful lock-in surface, and that competitive differentiation has to happen at the model-quality, sandbox-integration-quality, and ecosystem-depth layers. That is also implicitly an admission that the quickest way to extract value from the agent ecosystem is to plug in cleanly rather than to fight the current.
For anyone building with coding or tool-using agents, three immediate moves. First, if you are not already using AGENTS.md or an equivalent as your project's agent-configuration surface, adopt it now. The convention has settled and the cost of migration compounds the longer you stay on a custom system. Second, audit your MCP tool exposure. With OpenAI now officially supporting MCP, the server you built for Claude or for a third-party client will work natively in OpenAI's SDK, and vice versa. That removes a lot of friction from multi-provider setups. Third, choose your sandbox provider deliberately. Seven providers are now first-class in OpenAI's SDK, which means sandboxing is no longer a commodity decision. The providers differentiate on latency, language support, pricing, and security posture, and those differences will show up in your agent's runtime behavior. Cloudflare's Code Mode and E2B's Python-first stance are two examples of sandbox providers with distinct approaches. The meta-takeaway is that the agent standards have stabilized faster than most people expected. If you are still writing custom tool-calling scaffolds, custom sandbox plumbing, or custom instruction files, those are now more liability than asset.