OpenAI just made it a lot harder to ignore their agent platform. The company shipped two major upgrades to its Agents SDK: native sandbox execution and a model-native harness. If you run any kind of service business, this matters. It means building AI agents that actually run safely in production, for hours or days at a time, just got significantly easier.
This isn’t a research demo. It’s infrastructure. And it lands right in the middle of a heated race with Anthropic over who gets to be the default platform for production AI agents.
What happened
- Native sandbox execution is now built into the Agents SDK. This means your agent’s code runs in an isolated environment by default. If the agent does something unexpected (or dangerous), it can’t touch your main systems. Think of it like a padded room for AI code.
- A model-native harness now wraps the agent lifecycle. Instead of bolting on external monitoring tools, the SDK itself manages how the model interacts with files, tools, and long-running tasks. The harness is baked in, not an afterthought.
- Together, these updates let developers build agents that work across files and tools over extended periods, without the usual security headaches.
- OpenAI is positioning this directly against Anthropic’s managed agent offerings, which have been gaining ground with enterprise teams.
5 things service business operators should know about the Agents SDK update
- Sandbox execution changes the risk profile. Before this, running an AI agent in production meant building your own safety guardrails. Now, the SDK handles isolation natively. That’s one fewer thing to build and one fewer thing to break.
- Long-running agents are now a real option. Most agent setups today run for a few minutes at most. The combination of sandbox and harness means agents can operate over hours or days, handling multi-step workflows like document review, data processing, or client intake without someone babysitting them.
- You don’t need a huge engineering team to use this. The whole point of baking these features into the SDK is to lower the bar. A small dev team (or a good AI partner) can now deploy production-grade agents without reinventing security from scratch.
- This is a direct shot at Anthropic. Anthropic has been winning enterprise deals partly because their managed agent tools felt safer and more production-ready. OpenAI clearly wants that narrative back. Competition is good for you, the buyer.
- File and tool access is first-class now. Agents that need to read documents, call APIs, or interact with multiple tools can do so through the harness natively. No more hacking together fragile integrations just to let your agent open a PDF.
The hot take
This is OpenAI admitting that raw model intelligence was never enough. You can have the smartest model on the planet, but if it can’t run safely in a real business environment for more than five minutes, nobody’s going to trust it with actual work. The sandbox and harness aren’t flashy. They’re plumbing. And plumbing is exactly what the agent space has been missing. Anthropic figured this out first with their managed offerings, and OpenAI is now scrambling to close the gap. The winner of the agent platform war won’t be whoever has the best model. It’ll be whoever makes it easiest to deploy agents that don’t break things. This update is OpenAI’s clearest signal yet that they understand that.
The Agency OS play
If you run a service business (law firm, accounting practice, healthcare clinic, real estate brokerage, whatever), here’s what to do this week. First, identify one workflow that takes your team hours of repetitive work across multiple files or tools. Document review. Invoice processing. Client onboarding packets. Pick the one that makes people groan. That’s your candidate for a long-running agent.
Second, revisit any AI agent project you shelved because of security concerns. The sandbox execution feature changes the math on risk. If you previously decided “we can’t let an AI touch our client files without better isolation,” that isolation now comes out of the box. Dust off those plans and re-evaluate with the new SDK capabilities in mind.
Third, don’t lock yourself into one platform yet. OpenAI and Anthropic are both shipping agent infrastructure fast, and switching costs are still relatively low. Build your agent workflows in a way that keeps the core logic separate from the platform layer. Use abstraction where it’s cheap. The worst move right now is to go all-in on one vendor’s SDK without a plan for portability. The best move is to start building, start testing, and stay flexible while the two biggest players fight for your business.