OpenAI updates Agents SDK with sandboxing for enterprise deployments

OpenAI updated its Agents SDK on April 15, 2026, adding two capabilities aimed at enterprise deployments: sandboxed execution environments and an in-distribution harness for frontier model agents.

The sandboxing feature creates isolated workspaces where agents can only access explicitly approved files and tools. This addresses a practical concern that has slowed enterprise adoption of autonomous agents: the risk that an agent with broad permissions will take unintended actions on files, services, or systems outside its intended scope. With sandboxing, the agent’s operating environment is bounded from the start.

The in-distribution harness is a development and testing environment for agents running on OpenAI’s advanced models. It allows teams to deploy, test, and iterate on agents that perform multi-step, long-horizon tasks within a controlled workspace before moving to production. OpenAI’s product team described this as supporting “long-horizon agents” — tasks that take many steps and require sustained context across tool calls and intermediate outputs.

Both capabilities are initially available in Python, with TypeScript support coming later. Access is through the existing API at standard pricing.

For product managers building AI-powered features that involve autonomous agents, this update closes a gap that made enterprise deployment difficult. The ability to define a bounded workspace for an agent — controlling exactly which data and tools it can reach — makes it easier to scope and de-risk deployments that would otherwise require significant security review. It also simplifies the conversation with security and compliance teams, because the agent’s access surface can be defined explicitly rather than inferred from behavior.

The update reflects a broader shift in the agentic AI space toward infrastructure for controlled deployment rather than raw capability expansion.