Imitation learning
Imitation learning trains an agent from demonstrations of a task rather than from trial-and-error reward. For web agents the available demonstrations are interaction recordings, human sessions at scale, which makes behavioral data capture a training-data question and not only an analytics one.
How it works
The premise is that expert behavior encodes the policy you want: show the model enough examples of states paired with the actions competent users took, and it learns to act. Web-scale demonstration data has a shape problem that recording format determines: video of sessions is expensive to store and must be reverse-engineered into structure, while DOM event streams already are structure, with states, actions, and targets typed and timestamped. A corpus of recorded sessions filters into demonstrations (successful task completions, segmented by intent), and quality control becomes a data problem: separating expert behavior from confused wandering, which the same frustration signals used in analytics can score. The supervised core of the approach has its own entry at behavioral cloning; the wider framing of sessions as model fuel lives at training data.
Where you encounter it
Agent training literature, dataset papers for web tasks, and build-versus-buy debates about demonstration data.