OpenAI has updated its Agents SDK with a set of capabilities that reduce the technical overhead of building and deploying enterprise AI agents — provided those agents are being built within OpenAI’s own environment.

The April 15 update introduces three substantive additions. A model-native harness enables agents to work across files and tools on a local computer. A native execution sandbox provides a secure, isolated environment for running agent code. And sandbox-aware orchestration gives developers a managed layer for coordinating agent behaviour without having to build that infrastructure themselves.

The SDK itself has been positioning since early 2025 as the enterprise successor to OpenAI’s earlier “Swarm” multi-agent framework. This update extends that trajectory — but it also sharpens the trade-off at the centre of the product.

Forrester analyst William McKeon-White put it plainly: enterprise developers building inside the OpenAI ecosystem no longer need to manage their own agentic configurations and tooling; they can build on what OpenAI provides directly through ChatGPT. For those organisations, the update meaningfully lowers the barrier to agent deployment.

For larger organisations pursuing a provider-agnostic strategy, the picture is different. The new security and sandboxing capabilities are tightly coupled to OpenAI’s infrastructure. Enterprises building their own agent stacks — or working across multiple foundation model providers — will not benefit from this update in the same way.

The commercial logic is not hard to follow. Easier agent deployment means more token consumption and greater platform stickiness. OpenAI is making it structurally simpler to stay inside its ecosystem than to build outside it — a strategy that reflects its longer-term ambitions as much as its current product roadmap.

For business leaders evaluating their agentic AI infrastructure, the question this update raises is worth sitting with: how much of your agent stack do you want owned by a single provider?