MazeBench opens a 3D puzzle benchmark for coding agents with 200+ rooms
patience_cave · x · 2026-07-28
MazeBench: a public 3D puzzle benchmark for coding agents
MazeBench is an open-source local browser app for building persistent 3D puzzle worlds and evaluating coding agents inside the same JavaScript engine.
- The benchmark includes 200+ rooms and 100+ objectives (gems).
- It supports agent evaluation under different interfaces, including ASCII and Python-enabled modes.
- The posted leaderboard shows that giving agents Python can boost scores substantially, with GPT-5.6 Sol Max reaching 13%.
- The authors note that later puzzles are intentionally designed to be infeasible for A solvers, so current models have not yet reached those stages.
- They also use a novelty score to stop agents that start looping, keeping compute under control.
Related event: MazeBench Released as Open-Source 3D Spatial Reasoning Benchmark(4 posts)→
More from coding & agent
- Eve adds shadcn registry support for browser agent extensions — shadcn · 2026-07-28
- AI Tool Robin Reduces Dark Web Research to 30 Minutes — tom_doerr · 2026-07-28
- The real bottleneck in an Excel + LLM system is preserving business knowledge — Tired40s · 2026-07-28
- Four models tried the same game-building prompt, and Opus 5 looked finished — victor_explore · 2026-07-28
- Structured evaluation pipelines are becoming essential for production agents — blaizedsouza · 2026-07-28
- Retrying an agent step is not the same as safely resuming a 14-step workflow — msignificantdigit · 2026-07-28