Each WebArena task is modeled as a 4–7 agent pipeline:
Task Interpreter Agent
↓
[Site Browser Agents] (reddit / shopping / gitlab / wikipedia / map)
↓
Observation Aggregator Agent ← sensitive: captures raw page content
↓
Action Decision Agent
↓
Answer Verifier Agent
Total… See the full description on the dataset page:
https://huggingface.co/datasets/julee0323/webarena.
