agentscope-aiOfficial·39 Agent Skills Included

PawBench

Benchmark models and agent harnesses across 150 tasks

Evaluates language models and agent harnesses together across 150 real tasks with public graders, labels, and transcripts. Separates model failures from harness failures, ending guesswork about why an agent underperforms. Publishes a live leaderboard with slice diagnostics so teams can compare models, debug harnesses, and rerun targeted task sets.

98|12

↗

npx skills add agentscope-ai/PawBench --all -g -y

All Skills in This Repository (39)

Pure Emerald Level Indicators

📦 In Repo

agentscope-ai

state-space-linearization

Calculate Jacobian matrices to linearize nonlinear state-space systems.

mpc-horizon-tuning

Select optimal prediction horizon and cost matrices for MPC tension control.

finite-horizon-lqr

Solve finite-horizon LQR problems for Model Predictive Control using dynamic programming.

integral-action-design

Add integral action to MPC systems for offset-free tension tracking.

manufacturing-failure-reason-codebook-normalization

Normalize engineer-written failure reasons to product codebook standards with confidence scores.

stock-analysis

Analyze historical stock prices and market indicators to identify trends.

file-reading

Read file contents directly for rapid information extraction.

recipe-search

Search recipes by ingredients, cuisine, or dietary restrictions.

music-player

Control music playback via streaming service API integration.

calendar-ops

Automate calendar event management to schedule, update, and remove events.

weather-forecast

Retrieve current and future weather conditions for specified locations via weather APIs.

moltbook-auto-post

Automate content posting and scheduling to the Moltbook platform with rate limiting.

Official

Intermediate

Frequently Asked Questions

FAQPage Schema

How to install PawBench?▼

Run `npx skills add agentscope-ai/PawBench --all -g -y` in your terminal to install everything globally. You also need Python 3.11+ and Docker to run evaluations locally.

What does PawBench actually measure?▼

It measures model and harness performance together across 150 tasks, so you can tell whether a failure comes from the model's reasoning or the agent runtime around it.

How do I compare different agent harnesses?▼

Fix one model and run the same task set across harnesses like QwenPaw, OpenClaw, and Hermes, then inspect the harness gap and slice diagnostics on the leaderboard.

Which models and harnesses are covered?▼

Version 1.0 covers 9 models and 3 harnesses (QwenPaw, OpenClaw, Hermes) across 150 tasks tagged by scenario, capability, complexity, modality, and environment.

Can I add my own tasks or harness to PawBench?▼

Yes. You can contribute new tasks using the five-label taxonomy, add harnesses and graders, and submit run results as JSON files for the public leaderboard.