OpenAI’s Agents SDK update is about the boring layer that makes agents usable

OpenAI introduced new Agents SDK capabilities built around a model-native harness for work across files and tools, plus native sandbox execution for running agent work safely.

This is not the shiniest agent announcement. Good. The shiny part is often where demos live. The harness is where production either works or becomes a pile of clever prompts held together with incident reports.

Source credit: OpenAI's original source material.

The SDK is absorbing the messy primitives

OpenAI says the updated harness includes configurable memory, sandbox-aware orchestration, Codex-like filesystem tools, and integrations with patterns including MCP, skills, AGENTS.md, shell execution, and apply-patch style file edits.

That matters because useful agents need to inspect evidence, read and write files, run commands, use tools, and keep context across many steps. Teams have been building that plumbing themselves. Some did it well. Some created haunted automation.

• controlled workspaces for agent runs
• native sandbox execution
• filesystem and command-running tools
• memory and portable workspace manifests

Safety is architecture, not a slogan

The update supports bringing your own sandbox or using providers including Blaxel, Cloudflare, Daytona, E2B, Modal, Runloop, and Vercel. OpenAI also describes a Manifest abstraction for mounting files, defining output directories, and connecting storage providers such as S3, Google Cloud Storage, Azure Blob Storage, and Cloudflare R2.

The important security point is separation: keep credentials out of places where model-generated code executes, assume prompt injection and exfiltration attempts, and design the harness and compute layer accordingly. That is less fun than saying ‘agentic,’ but it is also how you avoid explaining to legal why the bot got creative.

The launch is generally available through the API with standard API pricing, with the new harness and sandbox capabilities launching first in Python and TypeScript support planned later.

The takeaway is straightforward: agent products are moving from clever wrappers to infrastructure choices. OpenAI wants its SDK to be the default path before every team builds the same risky scaffolding from scratch.