Benchmarks › Tasks › The error reader
Level 4 of 5 · Multi-step chain with rulesWorkspace task
The error reader
A cent is missing and the code looks right. Measures whether an agent runs things instead of skimming them — and whether it fixes the calculation or the test.
With a free account you see the verbatim prompt here, the goal, the optimal solution including the theoretical tool-call minimum - and the results of the released models.