OpenAI Relaunches Its Agents SDK — Native Sandbox, Codex-Like Harness, and Manifest for Long-Running Agents

On April 15, 2026, OpenAI shipped the next generation of its Agents SDK. The change is structural: the company is abandoning the "pure orchestrator" SDK and delivering a complete execution environment — sandbox, filesystem, model-native harness, and a Manifest abstraction to describe the workspace. Seven infrastructure providers are supported natively from day one: Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, Vercel. For the first time, a developer can take an OpenAI agent, plug it into their own E2B or Vercel account, and watch it run in an isolated container without writing a single YAML file.

What's Changing from v1

Agents SDK v1, released in late 2024, was a tool-call orchestrator. You wrote the tools, you wrote the sandbox, you wrote the state storage — the SDK just called the model and routed responses. In practice, most teams rewrote the 80% of code that should have been shared.

v2 absorbs that 80%. The built-in primitives — tool use via MCP, progressive disclosure via skills, custom instructions via AGENTS.md, shell execution, file edits with apply_patch — are now native. The SDK becomes what Codex was internally at OpenAI: a complete stack for running a coding agent on real code.

The second point — filesystem tools — matters most to developers. apply_patch is the tool Anthropic popularized with Claude Code 2: instead of asking the model to write a full file, you ask it to produce a minimal diff, then apply it with a robust parser. Fewer tokens, fewer regressions, atomic edits. OpenAI adopts the same primitive — a signal that the industry is converging on shared patterns.

Native Sandbox: 7 Providers Supported

This is the most concrete change. Running an autonomous agent that reads and writes code is a security problem before it's a model problem. Until now, every team had to pick a runtime (E2B for lightness, Modal for GPU-heavy work, Daytona for IDE integration) and write its own integration.

v2 ships with 7 official integrations:

• E2B and Modal for on-demand compute
• Blaxel and Runloop for long-running agents
• Daytona for dev-like environments
• Vercel and Cloudflare for serverless at the edge

You switch providers by changing a config. The Manifest describes the Docker image, environment variables, exposed tools — everything that used to be 200 lines of glue code.

For vibe coding startups, this is a relief. Building a Cursor-like or Replit-like platform meant building a homegrown sandbox. Now you can focus on UX and let the SDK handle execution.

Snapshotting: The Agent Picks Up Where It Left Off

The most underrated feature in this release: snapshot + rehydrate. The SDK captures container state — filesystem, environment variables, tool call history — and can restore it into a fresh container.

Practical consequences.

Long agents without timeouts. A cross-repo refactor that runs for 4 hours no longer dies when the runtime expires after 2. The SDK snapshots, relaunches a container, continues.

Debugging. You can replay exact state at step 37 of a 50-step run to understand where the agent failed. Git log, but for agents.

Parallelization. Start from the same snapshot to explore multiple solutions in parallel, then pick the best. That's the pattern Emergent Wingman uses for its autonomous agents out of India.

Manifest: AGENTS.md and Progressive Disclosure

Two conventions graduate from Codex into the standard SDK.

AGENTS.md — A markdown file at the repo root describing conventions, test commands, zones to avoid, and custom instructions. It's the OpenAI-Sourcegraph-JetBrains answer to Anthropic's CLAUDE.md, but with an open spec. The agent reads this file at startup and adjusts behavior.

Skills — Compressed prompt modules loaded on demand. Instead of injecting 40KB of context into every call, the SDK loads only an index, and the model requests relevant skills via a tool call. Same logic as Anthropic Claude Code's "progressive disclosure," but formalized in the OpenAI API.

Both primitives signal standardization. By late 2026, AGENTS.md and skills will likely be as universal as package.json or tsconfig.json.

What This Means for Competitors

Anthropic. Claude Code 2 had built a material lead in coding agents. Agents SDK v2 closes the infrastructure gap. The question becomes: whose model is better under the hood? With Opus 4.7 at 87.6% SWE-bench vs GPT-5.4 at 52%, the answer is still Anthropic — but OpenAI regains control over developer experience.

Cursor, Lovable, Replit, Windsurf. These platforms built their own harnesses. v2 could push them to adopt the OpenAI standard for long tasks, or alternatively to double down on proprietary IDEs to differentiate. Cursor at $9.2B has the resources to maintain its own stack; smaller players will migrate.

Vercel, Cloudflare, E2B. Being in the 7 launch providers is the best distribution imaginable. E2B and Modal were already standard in the AI ecosystem, but Blaxel, Runloop, and Daytona gain massive exposure. Expect accelerated funding rounds for those three within 60 days.

Microsoft / GitHub. Copilot already runs on OpenAI's SDK under the hood. v2 lets GitHub Copilot Workspace offer agents that run in GitHub's cloud (Codespaces) with snapshot support — the equivalent of a "cross-session undo."

The Limits

Python only at launch. The harness and sandbox aren't yet available in TypeScript. For an agentic ecosystem where TypeScript dominates (Next.js, Vercel AI SDK, Electron-based IDEs), that's a multi-month handicap. OpenAI promises TS support "soon" without a date.

Model lock-in. SDK v2 is optimized for GPT-5.4 and later. Running an agent with Claude or Gemini remains possible but forfeits the benefit of the model-native harness. Classic trade: deeper integration in exchange for lock-in.

Cost. Long-running agents with snapshotting multiply token consumption. A 4-hour run with 50 checkpoints can cost 30-50x a simple call. OpenAI hasn't yet offered differentiated pricing for agent mode.

Bottom line:

• Agents SDK v2 shipped April 15, 2026: native sandbox, filesystem tools, model-native harness
• 7 providers supported at launch: Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, Vercel
• apply_patch, shell, skills, AGENTS.md become native SDK primitives
• Snapshot + rehydrate: agents resume after container expiration or crash
• Manifest: a single abstraction describing the agent, its image, tools, and sandbox
• Python first, TypeScript support to follow
• Closes the infrastructure gap with Anthropic Claude Code, but doesn't compensate for the GPT-5.4 vs Opus 4.7 gap on SWE-bench

SDK v2 is OpenAI's answer to an uncomfortable truth: coding agents have become the killer app for LLMs, and OpenAI was losing the developer-experience war to Anthropic. The company regains control not by shipping a better model — that's for Spud / GPT-6 — but by delivering a complete, turnkey agentic stack that makes building a "Cursor-like" possible in a few hundred lines of code. It's the Android playbook: if you can't win on the end product, commoditize the platform beneath it. The result, 12 months from now, will be an ecosystem of coding agents where everyone has their own Cursor — and the only one who really wins is whoever provides the SDK.

Sources: OpenAI — The next evolution of the Agents SDK, TechCrunch — OpenAI updates its Agents SDK, Help Net Security — OpenAI updates Agents SDK, adds sandbox, Dataconomy — OpenAI Updates Agents SDK With Sandboxed Execution Tools.