Astra user reports 69 MazeBench gems with source-code access
The user says access to the source code removed the visual component, making the setup substantially easier than the official test. They limited the agent to one CPU and 16GB of memory.
TLDR
A user describes a MazeBench experiment with Codex that began with 5.6 sol and later resumed with Astra. The approach first used a fixed solver for individual rooms and another solver to stitch the solutions together, then moved to custom solvers tailored to specific rooms’ design and geometry.
The user reports that Astra reached 69 gems after a few more days of running, with the agent limited to one CPU and 16GB of memory. They say source-code access removed the visual component and made the setup substantially easier than the official test.
Combined views
36
1 Source, first seen 18d ago