# Alibaba Tongyi Launches General Agent Benchmark PawBench

- Category: Flash (快讯)
- Published: 2026-06-05T11:56:00.000Z
- Source date: 2026-06-05
- HTML: https://www.88z.ai/en/news/flash-20260605-9f769d57aee1b6d1
- Markdown: https://www.88z.ai/en/news/flash-20260605-9f769d57aee1b6d1.md
- Other language: https://www.88z.ai/news/flash-20260605-9f769d57aee1b6d1.md

Tongyi Lab has released PawBench, a general agent benchmark that jointly evaluates base models and execution harnesses for the first time. The benchmark includes numerous real-world tasks and test units across various model and harness combinations. Results show significant performance gaps between different harnesses, with score differences for the same model reaching up to 11.5 points, highlighting the critical impact of harness selection on agent performance.

## At a glance

- PawBench jointly evaluates base models and harnesses for the first time
- Score differences for the same model across harnesses reach up to 11.5 points

## The update

Tongyi Lab has officially launched PawBench, a general agent benchmark that innovatively incorporates both base models and execution harnesses into a unified joint evaluation system for the first time. PawBench v1.0 establishes a comprehensive testing matrix containing 150 real-world tasks and 4,050 test units, covering cross-combinations of nine models and three harnesses. The evaluation results reveal the significant impact of frameworks on performance. Data shows that performance gaps between different harnesses can reach up to 6.4 points. More critically, when the same model switches between different execution harnesses, the score difference can be as high as 11.5 points. This finding indicates that when assessing agent capabilities, attention should not be limited to the model itself, as the choice of execution harness plays a decisive role in final performance.

## Sources

- [通义实验室](https://mp.weixin.qq.com/s/Q1fa3KwT63HBOF2fmWKzlg) (2026-06-05)

This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.