Integrate AI agents directly into your Windmill flows with comprehensive support for all major AI providers, multimodal capabilities, streaming responses, tool integration, and structured outputs. AI agents can process text and images, call Windmill scripts as tools, stream responses in real-time, and output structured data following JSON schemas.

Extended streaming capabilities to flows in addition to scripts, enabling real-time streaming for AI agents and other flow steps. Added new SSE endpoints to start and listen to jobs with comprehensive streaming support.

We introduce native result streaming in 1.518.0. Returns a stream (any AsyncGenerator or iter compatible object) as a result OR to stream text before the result is fully returned. It works with Typescript (Bun, Deno, Nativets), Python and is compatible with agent workers. If not returning the stream directly, we introduce 2 new functions on our SDK: wmill.streamResult(stream) (TS) and wmill.stream_result(stream) (Python), to do it mid-script. It's made with LLM action in mind because many scripts nowadays interact with LLM that have streaming response that can be returned in a streaming manner as a result. We will progressively refactor all our relevant hub scripts that are compatible with streaming to leverage this new capability. The stream only exists while the job is in the queue. Afterwards, the full stream becomes the result (or added as the field "wm_stream" if there is already a result).