OpenAI Updates the Agents SDK (April 15, 2026): Native Sandboxes, MCP Tools, and Codex-Like File Workflows

OpenAI published "The next evolution of the Agents SDK" on April 15, 2026, describing a meaningful step toward more reliable, long-running developer agents: a model-native harness plus native sandbox execution for running tool-using work safely.

TL;DR

• A more standardized agent harness designed around how frontier models actually operate.
• Native sandbox execution so agents can run commands and handle files in controlled environments.
• Clear primitives for production agent stacks: MCP tool use, AGENTS.md instructions, and Codex-like filesystem tools.

What Shipped (Verified)

According to OpenAI’s announcement, the updated Agents SDK includes:

• A model-native harness to let agents work across files and tools on a computer.
• Native sandbox execution so developers can run agent work safely in controlled environments.
• Configurable memory and sandbox-aware orchestration.
• Codex-like filesystem tools for file-centered workflows.
• Standardized integrations and primitives, including:
  • Tool use via MCP
  • Custom instructions via AGENTS.md
  • Code execution via a shell tool
  • File edits via an apply-patch style tool

OpenAI also describes a manifest abstraction for describing an agent workspace (mounting local files, output directories, and pulling data from common object stores like AWS S3, Google Cloud Storage, Azure Blob Storage, and Cloudflare R2).

Sandboxes: Built-In Providers + Bring Your Own

For teams that don’t want to build sandbox infra from scratch, OpenAI lists built-in support for multiple sandbox providers:

• Blaxel
• Cloudflare
• Daytona
• E2B
• Modal
• Runloop
• Vercel

OpenAI also explicitly supports "bring your own sandbox".

Why This Matters for Developer Tooling

Most agent frameworks start as a thin wrapper around model calls, then grow ad-hoc tool execution, file handling, and state. OpenAI’s positioning here is the opposite: start with a harness + sandbox that makes long-horizon work safer and more reproducible, then layer product behavior on top.

Two practical implications for teams building AI devtools:

1. Sandbox-first becomes default: if your agent can run shell commands and edit files, you probably want those actions isolated.
2. MCP standardization keeps spreading: MCP is being treated as a first-class interface for tool use, not just an add-on.

Source

• OpenAI announcement: https://openai.com/index/the-next-evolution-of-the-agents-sdk/