What problem do sandboxing and durable execution solve?
On 15 April 2026, OpenAI introduced a model-native harness and sandbox execution for the Agents SDK. Agents can work with files and tools in a controlled environment and use snapshots for longer tasks.
These capabilities separate repeatable, recoverable execution from an application's main logic, allowing long tasks to resume from explicit state rather than remain an opaque session.
How can file and tool use be made safer?
File and tool work increase attack surface and error risk. Separating execution can reduce unnecessary access to credentials and primary systems.
Separation is only one control; tool permissions, network limits, traces and human approval still matter. Organizations need to know which files, connectors and outputs an agent could reach.
A practical pattern for enterprise rollout
Organizations should use minimum access, ephemeral environments, time limits and stop mechanisms. Durable state should remain reviewable and avoid unintended sensitive-data retention.
Start with one bounded task, set a stop criterion and approver, then review execution traces against failure examples. This turns a new capability into a dependable operating control.