Show HN: Action Engine — An API/Agent Buildkit Putting Flexibility First (actionengine.dev)

Google DeepMind has released Action Engine, an open-source (Apache 2.0) toolkit for building multimodal, streaming APIs and UIs that emphasizes modularity and flexibility over rigid “agentic” frameworks. Instead of imposing heavy abstractions, Action Engine supplies a small set of clear primitives—actions (named executable units with i/o schemas and lifecycle hooks) and async nodes (channel-like streams with blocking-read-with-timeout semantics and nonblocking writes). These building blocks let you stream text, images, audio and custom types bidirectionally between clients, servers, edge devices or peer-to-peer WebRTC transports, and you can pick storage/transport per-stream (in-memory, Redis Streams, or your own backend). For AI/ML developers this matters because it addresses real needs for stateful, long-running, dynamic workflows in multimodal and agentic systems without locking you into a single runtime or opinionated graph executor. Action Engine supports incremental, experimental composition (JS/Python/C++ examples, FastAPI integration), easier handling of streaming chunks and intermediate named streams (logs, thoughts, multimodal outputs), and flexible deployment options. The project aims to accelerate prototyping and interoperability across diverse applications while leaving performance, safety, and persistence choices to the developer—making it a pragmatic common-building-block for next‑generation AI infrastructure. Source and demos are on github.com/google-deepmind/actionengine.