OpenAI is expanding the Responses API with features aimed directly at long-running AI agents. The update focuses on a practical problem: how to keep agent work moving over extended sessions without forcing every workflow, dependency, and instruction into a single overloaded prompt.
The three additions are server-side compression, controlled internet access for OpenAI-hosted containers, and "skills" that package reusable instructions, scripts, and files. Together, they point to a more structured way to build agents that can keep context, run code, and reuse specialized workflows when needed.
Why longer agent sessions need new support
Long-running AI agents create a different set of demands from short prompt-and-response interactions. They may need to hold onto earlier work, continue across many steps, and execute actions inside a container while still staying within context limits.
The first major feature in the update is server-side compression. According to the source article, it is designed to let agent sessions continue for hours without pushing past context limits. That matters because context is one of the core constraints for any agent that has to remember what it has already done.
Instead of leaving developers to manage every piece of accumulated session state manually, server-side compression gives the Responses API a built-in way to reduce context pressure. The article does not describe the internal mechanics of the compression, but the purpose is clear: longer sessions should be able to continue without being derailed by context growth.
Containers get controlled internet access
The second major addition is controlled internet access for OpenAI-hosted containers. The source article says this allows those containers to install libraries and run scripts.
That is a meaningful change for agent workflows that depend on code execution. If an agent needs a library in order to complete a task, the hosted container can have a path to install it rather than being limited to whatever is already present.
The word "controlled" is important. The article does not describe an open-ended browsing environment; it describes internet access for OpenAI-hosted containers in a controlled form. The practical takeaway is that OpenAI is giving hosted agent environments more ability to prepare and execute work, while still framing that access as governed rather than unrestricted.
For developers using the Responses API, this can make hosted containers more useful as execution environments. An agent can run scripts, bring in needed libraries, and carry out longer pieces of work without requiring every dependency to be preloaded into the environment.
Skills create reusable agent building blocks
The third feature is "skills." The source article describes them as reusable bundles of instructions, scripts, and files that agents can load and execute when needed.
Skills sit between system prompts and tools. That position is important because it gives developers another layer for organizing agent behavior. A system prompt can stay focused on durable rules and role definition, while a skill can hold a specific workflow that is only relevant in certain situations.
OpenAI’s approach, as described in the source, avoids the need to put long workflow instructions into every prompt. Instead, developers can package those workflows as bundles. The agent can pull in the relevant bundle on demand, keeping the main prompt cleaner and more focused.
The article also says skills ship as ZIP files, support versioning, and work in both hosted and local containers through the API. That makes them portable across different execution setups and gives teams a way to manage changes over time.
How developers are expected to use skills
OpenAI recommends building skills like small command-line programs and pinning specific versions in production. That guidance suggests skills are meant to be treated as operational components, not just prompt snippets.
A skill can contain multiple kinds of material:
- Instructions that tell the agent how to carry out a workflow
- Scripts that the agent can run when the workflow requires execution
- Files that support the task the skill is built to handle
Versioning also matters for production use. If a skill changes, the behavior of the agent that uses it may change as well. Pinning specific versions gives developers a way to keep production behavior tied to a known bundle rather than whatever version happens to be latest.
The source article does not provide examples of particular skills, but the architectural role is clear. Skills are meant to make reusable agent capabilities easier to package, distribute, and invoke without expanding every prompt.
What the Responses API update signals
This update moves the Responses API further toward supporting agents as long-running software systems. The emphasis is not only on model output, but also on session continuity, execution environments, dependency handling, and reusable workflow packaging.
Server-side compression addresses the context problem. Controlled internet access makes hosted containers more capable. Skills give developers a cleaner way to separate reusable procedures from general instructions and tools.
For teams building AI agents, the broader implication is organizational. Agent behavior can be split into more manageable pieces: the prompt defines the role, tools provide capabilities, containers execute work, and skills package task-specific workflows. That structure can make agent systems easier to maintain as they grow more complex.
The Responses API changes do not remove the need for careful engineering. Developers still need to decide what belongs in a prompt, what should be a tool, and what is better packaged as a skill. But the new features give them more specific primitives for building agents that are expected to work over longer sessions.