OpenAI released a significant update to its Agents SDK on April 15, 2026, introducing sandboxed execution environments, a model-native harness, and support for long-horizon tasks. For any organisation building or evaluating AI agent systems, the update marks a genuine step forward in making agents safe enough for production deployment.

What Changed

The core additions are two architectural pieces that have been missing from most agent frameworks.

Sandbox execution isolates agents inside controlled computer environments. Rather than letting an agent run freely across a system, the sandbox limits what files and tools it can access to a defined workspace. If an agent makes a mistake or behaves unexpectedly, the blast radius is contained. This addresses one of the biggest objections enterprises raise when considering agents: the fear of a runaway process doing something it shouldn’t.

The model-native harness is the scaffolding that connects an agent to its workspace. It includes configurable memory, sandbox-aware orchestration, and filesystem tools modelled on how Codex approached code interaction. The practical effect is that agents can now handle multi-step tasks across files and tools without developers having to build all that plumbing from scratch.

Together, these features support what OpenAI calls “long-horizon” tasks: operations that require many steps, decisions, and tool calls spread over time, rather than a single prompt-and-response exchange.

Additional capabilities in the update include provider-agnostic support for over 100 LLMs (not just OpenAI’s own models), subagents (agents that can spawn and direct other agents), and a code mode that lets agents write and execute code as part of their workflow. The new capabilities are available through the existing API at standard pricing with no separate tier. Python support launched first, with TypeScript coming later.

Why Sandboxing Is the Feature That Actually Matters

It’s easy to focus on the capability upgrades: long-horizon tasks, 100+ LLM support, subagents. But for businesses considering their first serious agent deployment, the sandbox is the more important development.

Most organisations have stalled on AI agents not because the technology doesn’t work, but because the risk profile felt unacceptable. Agents that can read files, write code, send emails, and interact with databases need guardrails before they go anywhere near production. The sandbox provides a structural answer to that concern rather than a procedural one.

This shifts the conversation from “can we trust this agent?” to “is this workspace configured correctly?” That’s a much more manageable question for IT teams and security functions.

The Provider-Agnostic Move

OpenAI supporting 100+ LLMs inside its own SDK signals a strategic shift. A developer can now build on OpenAI’s agent infrastructure while routing tasks to Anthropic’s Claude, Google’s Gemini, or open-source models as needed. For enterprises with multi-vendor AI strategies (which is most of them at this point), this removes a friction point that previously forced a choice between using OpenAI’s ecosystem or building your own.

What This Means for Business

If you’re a business owner evaluating whether to invest in AI agents, this update is a signal rather than a direct product for you. It tells you that the underlying infrastructure for agent deployment is maturing fast.

Three things to take away:

Agents are ready for complex work. Long-horizon task support means agents can now handle workflows that unfold over hours or days, not just instant lookups or single-step automations. Payroll reconciliation, procurement review, client onboarding: all of these become realistic targets.

Safety is becoming standardised. Sandboxing is now a built-in feature, not a custom build. As more frameworks adopt this approach, the compliance and security questions around agents get easier to answer. That removes one of the main barriers to budget approval.

The vendor landscape is consolidating around infrastructure. OpenAI building a harness that works with any LLM is a bet that developers will prefer one framework over many. For businesses, that means fewer integration headaches and more predictable vendor relationships.

The gap between “AI agents as a pilot project” and “AI agents as operational infrastructure” is closing. For organisations still in evaluation mode, the window to start building institutional knowledge about agent deployment is narrowing.

---

AI agents are moving from experimental to operational. The new SDK makes it practical for engineering teams to deploy agents with real safety guarantees, not just aspirational policies. If your business is still waiting for “the right time” to explore agents, that time is now.

For a deeper walkthrough of tools like this and how they fit together, the free Working With Claude field guide covers the ecosystem end to end. Get the guide.