Agents with Tools Fail Together

Abstract

One public asset for the axis the earlier notes set aside: 175 software agents, each a model wrapped in its own tools and workflow, scored on the same 500 SWE-bench Verified bugs. The seven best carry 1.4 effective independent votes, not seven; a majority of them resolves 79.4% of bugs against 79.2% for the best one alone, zero inside the interval; a perfect selector over the seven would reach 89.0%, ten points of coverage that no judge in the earlier notes could collect. Forty-one models in one scaffold and one model in eleven scaffolds give the same shape. Twenty-six bugs are resolved by none of the 175. On τ-bench's released logs one agent rerun four times is worth 1.4–1.9 independent runs, any-of-four adds 14–25 points over one run, and four runs of the best model cover as much as four models once. Pre-registered against the text-corpus predictions, zero inference, results only, no trajectories read.