AIエージェント
ClawBench
An open benchmark for browser AI agents on everyday tasks.
★ 975 スターApache-2.0
ガイドを見る →WorkArenaは、ServiceNowが公開しているオープンソースのAIエージェントプロジェクトです（GitHub: ServiceNow/WorkArena）。GitHubスターは275、フォークは40で、NOASSERTIONライセンスの下で公開されています。GitHubでの最終更新は2026年9月28日です。
WorkArena: ServiceNow's benchmark for how well web agents handle everyday knowledge-work tasks.
Decide whether you will evaluate on the atomic L1 tasks or the compositional WorkArena++ (L2/L3) tasks.
Request access to the gated WorkArena instances dataset on Hugging Face and authenticate your machine.
Install browsergym-workarena and Playwright into your BrowserGym environment.
Run the demo to see tasks solved by the oracle cheat functions, then wire your agent through env.step and validate its results.
WorkArena-L1 contains tens of thousands of unique instances across 33 tasks covering the main ServiceNow interface components.
Hundreds of composed tasks combine atomic skills into realistic use cases that test planning, reasoning, and memory.
Tasks run on actual ServiceNow instances rather than toy websites, reflecting real knowledge-work software.
The preferred evaluation path runs parallel experiments through BrowserGym with a unified public leaderboard.
Each task ships cheat (oracle) and validate functions, so you can see a task solved and then score your own agent against the same checks.
The benchmark is documented in ICML 2024 and NeurIPS 2024 papers linked from the repository.
Most evaluators install the package from PyPI into a BrowserGym environment.
git clone https://github.com/ServiceNow/WorkArena.git
cd WorkArenaFirst gain access to the gated instances dataset on Hugging Face and authenticate, as the README requires.
pip install browsergym-workarena
playwright installInstantiate tasks through BrowserGym's BrowserEnv and drive them with your agent via env.step, validating with the task's validate function. # see README live demo code
# see the README Live Demo section for the full BrowserEnv exampleIt is maintained by ServiceNow's research team.
The package is openly available, but the catalog metadata lists its license as NOASSERTION, and running it requires approved access to the gated ServiceNow instances dataset, so read the repository terms first.
No LLM key is needed for the benchmark itself; the agent you evaluate brings its own model access, and you need Hugging Face authentication for the gated instances.
WebArena tests agents on self-hosted web apps like shopping and forums; WorkArena tests them inside ServiceNow on enterprise knowledge-work tasks such as forms, catalogs, and dashboards.
Get approved for the instances dataset, pip install browsergym-workarena, install Playwright, and run the README's demo script through BrowserGym.
An open benchmark for browser AI agents on everyday tasks.
OSWorld — benchmark multimodal agents on open-ended tasks in real computer environments.
Simple desktop AI assistant with knowledge bases, agents, and web search.
An open-source, privacy-first, self-hosted knowledge workspace where humans and AI agents work together 开源、隐私优先、自托管的知识工作空间，让人与智能体在此协作
A benchmark testing whether agents can solve long-horizon coding tasks.
Benchmark computer-use agents on long, real-world desktop tasks.
LibreAIカタログに掲載されたAIエージェントプロジェクトの中で最もスターの多いもの — WorkArenaが合わなかった場合に検討したい、現実的な候補リストです。
A personal AI that really does things — on any OS, with your data staying yours.
An agentic skills framework & software development methodology that works.
Agent harness optimization system with skills, memory and security for Claude Code, Codex and more
The agent that grows with you — a Python framework for adaptive AI agents.
DeepSeek's plugin-based agent harness where everything is a plugin.
カタログの中でWorkArenaとスター数が近いAIエージェントプロジェクト — 前後の棚もそのまま見て回れます。
Benchmark computer-use agents on long, real-world desktop tasks.
A complete evaluation toolkit for AI agents.
Agentic memory for LLM agents, from a published research project.
An MCP-native AI coding agent and runtime in a single binary.
Safety testing for agents, right in pytest.
A benchmark for comparing AI agent memory systems head-to-head.