Action space
An action space is the set of operations an agent can perform on its environment; for web agents, typically click, type, select, scroll, navigate, and wait. The recorded interaction vocabulary of session replay maps onto the same set, which is what makes human sessions legible as agent demonstrations.
How it works
Action space design trades expressiveness against reliability: a small, coarse action set is easier for models to use correctly but cannot express drag interactions or precise text editing, while richer spaces multiply failure modes. The web's de facto action space converged on a handful of primitives, and rrweb's incremental sources (mouse interaction, input, scroll, viewport, media) record human behavior in nearly the same vocabulary. That alignment carries practical weight in both directions: recorded human sessions translate into state-action sequences a training pipeline can consume (see behavioral cloning), and recorded agent runs can be validated action by action against what the interface state permitted at each step.
Where you encounter it
Agent framework APIs, RL formulations of web tasks, and the schema design of trajectory datasets.