OpenAI announced they are extending the Responses API to make it easier for developer to build agentic workflows, adding support for a shell tool, a built-in agent execution loop, a hosted container workspace, context compaction, and reusable agent skills.
With the new capabilities, OpenAI says, developers using the Responses API will not need to build their own execution environments to safely and reliably execute real-world tasks. Instead, they can rely on a built-in infrastructure designed to handle the practical challenges that all agent developers face, such as managing intermediate files, optimizing prompt usage, ensuring safe network access, and handling timeouts and retries.
At the heart of this new design is the agent execution loop. Instead of producing a final answer immediately, the model can "propose" an action, such as running a command, querying some data, or fetching something from the internet. That action is executed in a controlled environment and the result is fed back into the model. This cycle continues iteratively until the task is complete.
The execution loop can leverage the new Shell tool to interact with a computer via the command line, allowing it to carry out tasks using familiar Unix utilities like grep, curl, and awk, all available out of the box, and other programs.
Compared to our existing code interpreter, which only executes Python, the shell tool enables a much wider range of use cases, like running Go or Java programs or starting a NodeJS server. This flexibility lets the model fulfill complex agentic tasks.
It is important to understand that the model can only propose using a tool, not run it itself.
In addition to the Shell tool, OpenAI is also bundling a containerized execution environment where files and databases can live, with network access safely managed through policy controls. These tools are designed to reduce the need for packing all input directly into prompt context. Instead, the model can access specific files using shell commands to parse or transform the available information, run a database query to pull in only the relevant rows, or fetch the data directly from the network.
To make network access safe, all outbound traffic is routed through a centralized policy layer that enforces allow-lists and access controls while keeping traffic observable. Credentials are not stored inside the container and remain invisible to the model, which only sees placeholders that are replaced in the external layer.
Another key idea introduced with the new Responses API is "skills", which make it easier to define complex, repeatable tasks in which shell commands are executed according to a predefined pattern.
Agent skills package those patterns into reusable, composable building blocks. Concretely, a skill is a folder bundle that includes
SKILL.md(containing metadata and instructions) plus any supporting resources, such as API specs and UI assets
The new Responses API also addresses the challenge of managing context size for long-running tasks, which inevitably tend to exceed context limits. To handle this, the system uses compaction, compressing previous steps into a shorter representation while preserving the important information, similarly to what Codex does. This allows the agent to continue working over many iterations without exceeding token limits.
According to OpenAI, by combining the orchestration capabilities provided by the Responses API, along with shell tools, container-based runtime context, skills, and compaction, developers can build agents capable of executing long-running tasks from a single prompt. For the full detail, be sure to check the original announcement.