HAL dashboard tracks agent reliability gaps

This is a smart push against leaderboard theater: agent systems do not become production-ready just because pass rates inch upward. For AI developers, the real question is whether an agent fails consistently, predictably, and safely under pressure.

• –The accompanying paper introduces 12 reliability metrics and finds that capability gains have produced only small improvements in reliability across 14 evaluated models.
• –The dashboard makes those failure modes visible across benchmarks like GAIA and τ-bench, which is more actionable than a single top-line score.
• –HAL frames reliability as engineering infrastructure, not just academic critique, giving teams a clearer way to compare agents before deploying them.
• –Open, third-party evaluation here matters because it pressures vendors to show how agents degrade and fail, not just how often they eventually succeed.