01
Available now
Offline reference website
A stateful, isolated server the agent can inspect only through interaction.
Agent website reconstruction benchmark
Can an agent rebuild a website it can only explore?
WebsiteBench gives an agent an isolated offline website—not its source code. The agent must interact with it, infer its behavior, and produce a new runnable clone. This viewer makes the input website, the generated clone, and every score inspectable.
Dataset construction · active
Agent experiments · not started
Amazon Shopping · worked exampleOpen evidence →
02
Reserved
Agent exploration
Browser actions, visited states, tool traces, and resource use will appear here.
03
Reserved
Generated clone
The new runnable server produced by each evaluated agent.
04
Reserved
Multimodal scoring
Screenshots, interaction replay, terminal states, robustness, and efficiency.
Dataset item 01 · accepted
One complete offline website, fully inspectable
Amazon is the current worked example. The dataset is intentionally sparse while additional website categories are still being built.
Standard offline reference · accepted
Purpose
A bounded shopping environment spanning discovery, product configuration, identity, checkout, orders, returns, reviews, and recovery.
Success journeyFailure journeyRecovery journey
Calibration ≠ agent score
The source-to-offline comparisons below certify the benchmark input. They are not a model result.
Corpus roadmap
More website categories will be added here
0102 03 04
Amazon Shopping
Accepted
Shopping Commerce
Future website
ReservedCategory and scope not assigned
Future website
ReservedCategory and scope not assigned
Future website
ReservedCategory and scope not assigned
Future experiment output
Every agent run will stay explainable
No agent reconstruction has been evaluated yet.
Model names, generated clones, replays, and scores remain blank until real experiment reports are added.