WorkThrough-Bench: Evaluating Long-Horizon AI Agent Productivity Work

openreview.net38 min read

Researchers introduce WorkThrough-Bench, a benchmark of 86 long-horizon tasks across dozens of industries designed to evaluate AI agents on deep research, deliverable creation, and application building. Evaluating six frontier models and two agent harnesses, they find that while agents complete an average of 68.0% of required work, task pass rates remain below 11% due to self-contented completion, where agents judge tasks against their own implementation rather than the original contract. Introducing static feedback that redirects verification toward the original requirements improves the mean task pass rate by 26.7%. Open article →

More on this topic