The biggest problem with AI agents isn't whether they can work, but whether you dare let them
Over the past year, stories of AI agents "going rogue" have been common: sending emails they shouldn't have, deleting files they shouldn't have, turning a small task into a cascade of side effects.
This isn't about models not being smart enough—it's about execution environments lacking security boundaries.
OpenAI's April 15 update to the Agents SDK directly addresses this issue: it introduces a sandbox mechanism that lets agents run in an isolated computing environment, unable to access files and tools they shouldn't touch.
How the sandbox works
Simply put: the agent is confined to a "workspace" where it can only access explicitly permitted files and code paths, and cannot interact with the rest of the host system.
The SDK now supports multiple third-party sandbox providers. Enterprises can bring their own infrastructure or directly integrate with the following platforms:
| Sandbox Provider | Features |
|---|---|
| Cloudflare | Edge computing, low latency |
| Vercel | Frontend-friendly, function sandbox |
| E2B | Focused on AI code execution |
| Modal | Elastic compute |
| Daytona | Development environment isolation |
| Runloop | Enterprise-grade workflows |
| Blaxel | Dedicated to AI agents |
Which one you choose depends on your existing tech stack; most teams should find a match.
What else is new
Beyond the sandbox, this update also brings:
- Model-native Harness: Supports agents in performing cross-step, long-chain tasks across files, code, and tools, moving beyond single-turn interactions
- Sub-agent support (roadmap): Future capability to let agents split into multiple sub-tasks for concurrent execution
- Code mode (roadmap): A special agent behavior pattern better suited for code execution
Karan Sharma from OpenAI's product team said they want developers to "build long-horizon agents using our harness and their own infrastructure."
What are the limitations
Currently, only the Python version is available. TypeScript support is on the roadmap but no timeline has been given.
This is an inconvenience for frontend-led teams, but less of an issue for backend and data engineering teams.
Pricing follows standard API billing with no additional fees.
Why this matters
In the agent race, Anthropic has Managed Agents, Microsoft has the AutoGen framework, and OpenAI's previous SDK was relatively basic.
With this update, OpenAI's agent infrastructure finally has basic enterprise-grade security capabilities—the sandbox mechanism is a prerequisite for production deployment; without it, enterprises simply wouldn't dare put agents into critical workflows.
The timing is also interesting: right as Anthropic pushes Claude-hosted agents to enterprises, this OpenAI move feels more like a defensive update than an offensive one. But for developers, having both companies seriously building agent infrastructure is ultimately a good thing.
Sources: OpenAI updates its Agents SDK to help enterprises build safer, CocoLoop, more capable agents (TechCrunch)