OpenAI launches the Responses API and Agents SDK

Background

Developers calling OpenAI's models through its API had, by early 2025, two broad styles to choose from. Chat Completions was the simple path: send a conversation, get a reply. The Assistants API added the machinery of tool use — the ability for a model to reach beyond text, keep working across steps, and draw on outside resources ^[5]. Building an agent on top of either meant assembling the pieces yourself, and seeing what an agent actually did meant keeping your own record of it.

What launched

On 11 March 2025, OpenAI introduced the Responses API together with an open-source Agents SDK ^[1]. OpenAI described the Responses API as combining the simplicity of Chat Completions with the tool-use capabilities of the Assistants API ^[5]. In other words, it was presented not as a third, unrelated interface but as a meeting point of the two earlier ones: the short request-and-reply shape developers already knew, carrying the tool machinery that had previously lived in the Assistants line.

Key details

The launch was bundled rather than piecemeal. It included built-in web search, file search and computer use tools, plus integrated observability for tracing agent workflows ^[5]. Web search and file search let a model pull in material from outside its own parameters; computer use let it act on a screen. Tracing addresses a different problem: an agent that takes many steps is hard to debug when only the final answer is visible, so the platform records the path through those steps.

The computer-use tool had a specific origin. The Computer-Using Agent model behind Operator, OpenAI's computer-use agent, became available to developers as computer-use-preview through the Responses API ^[4]. A capability first shown inside a product was thus exposed as a developer tool.

OpenAI also framed the API around the size of the task rather than the size of the request. It said a single Responses API call could handle increasingly complex tasks using multiple tools ^[5] — the unit of work being a job with several tool calls inside it, not one exchange.

What followed

The Responses API did not remain a standalone release. AgentKit, announced in October 2025, was built on top of it ^[2], which placed the API in the position of a foundation layer for later agent tooling.

The older interface moved the other way. The Assistants API was scheduled to shut down on 26 August 2026 ^[3], giving developers a migration window with a fixed end date.

Why it matters

The launch changes what an API call is for. Chat-style interfaces were designed around producing a reply; the Responses API was designed around completing a task with tools, and around being able to see how it was done. Bundling search, file access and computer use as built-in tools — with tracing in the same package — moved work that developers had previously done in their own glue code into the platform itself ^[5]. The Agents SDK, released as open source alongside it, covered the orchestration half of the same shift ^[1].

For anyone learning how agent systems are assembled, the episode is a compact example of a wider pattern: capabilities first demonstrated in a single product get generalised into interfaces, and the interfaces in turn become the base that later tools are built on ^[2].