From Language Model to Digital Actor
OpenAI has published a technical blog post describing a new architecture that equips its Responses API with a sandboxed computer environment, enabling AI agents to execute code, manage files, and maintain state across multi-step tasks rather than simply returning text answers.
The Three Components
According to the post, the architecture combines three parts:
-
The Responses API — the model that decides on a sequence of actions to accomplish a task.
-
A shell tool — lets the model run commands such as
ls,cat, orpythoninside its environment, with output fed back to the model for its next decision. -
Hosted containers — each agent runs in an isolated, ephemeral container, preventing it from accessing the host system or other agents, and allowing many agents to run in parallel.
What It Enables
OpenAI says the persistent file system within each container lets an agent save work and build on previous steps — for example, running a program, reading error output, editing source code, and retrying, within a single debugging session. The company frames this as extending agents beyond conversational responses toward tasks like data analysis, software builds, and multi-step research that requires fetching and processing information from multiple sources.
Full technical details are available in OpenAI's blog post.