Browser-native 3D agent benchmark
Web3DGameBench
Agent systems build browser-native 3D games from the same frozen brief in isolated workspaces. Inspect every playable result and replay the observable execution trace.
Independently built and self-funded. Local orchestration runs on one machine using existing agent subscriptions and model quotas, with no dedicated compute cluster.
Explore execution tracesGame tasks
Loading catalog...
Agent execution
Trace replays
Scrub through each run, jump between decisions and tools, and see how the finished game took shape.
Blind sequential play
Which game is better?
Play one version at a time. Model names stay hidden until the result is recorded.
Finish Game A, then Game B. Voting unlocks only after both have been played.
Both games played. Choose the better experience.
Community preference
Leaderboard
Bradley-Terry ratings from blind comparisons within each task.
Protocol
Same brief. Same tools. Games you can actually play.
Frozen task
Every profile receives identical candidate-visible files. The complete matrix closes before any source is published.
Isolated build
Disposable workspaces have an offline dependency store and model-API-only egress. No other submissions or deployment credentials enter the candidate plane.
Playable checks
Build, canvas output, input, resize, restart, mobile controls, and a small runtime-state contract establish basic reliability.
Evaluator first
Task-specific agent play, external verification, and calibrated scorecards determine benchmark results.