OpenAI launches Responses API and agent-building tools

OpenAI launched the Responses API, a new stateful interface for developers building agents, alongside an open-source Agents SDK. The API bundled three built-in tools that had previously required custom integration or existed only inside OpenAI’s own products: web search, file search over uploaded documents, and computer use — letting a model take actions inside a virtual browser or desktop environment, the same capability underlying OpenAI’s Operator product.

The release combined elements of the company’s earlier Chat Completions and Assistants APIs into a single interface intended to let third-party developers build the kind of autonomous, multi-step workflows OpenAI’s own agents, such as Deep Research, already used, without developers having to write their own orchestration logic to chain tool calls together. OpenAI positioned the older Assistants API for eventual retirement, though it did not formally deprecate it until several months later.

The launch was part of a broader push across the industry, in the same weeks as Anthropic’s own agent-oriented protocol work and Google’s Gemini Robotics announcement, to standardise the plumbing around agentic systems — search, memory, tool use, computer control — as reusable building blocks rather than bespoke integrations each developer had to assemble. It set the infrastructure OpenAI built on later that year with Codex and other coding-agent products.

Developer reaction was mixed in the following months: some engineers welcomed the built-in tools as a genuine simplification, while others found the transition from Chat Completions or Assistants confusing enough that an OpenAI staff member later posted a public thread attempting to clarify how the new interface was meant to be used. That confusion did not stop adoption — within weeks OpenAI was extending the API further, adding support for the Model Context Protocol and native image generation — but it underlined a recurring pattern in the agent tooling race: capability shipped faster than documentation or developer understanding could keep pace with it.