InternLM/WildClawBench
An in-the-wild benchmark for AI agents in the OpenClaw Environment.
What it solves
Most AI agent benchmarks focus on isolated capabilities like parsing JSON or calling single functions. WildClawBench addresses the need for testing real-world, end-to-end agency by dropping agents into live environments to perform complex, multi-step tasks that require autonomous planning and error recovery.
How it works
The benchmark uses 60 hand-built tasks across six categories (Productivity, Code, Social, Search, Creative, and Safety) executed within a live OpenClaw environment. It utilizes Docker containers to ensure each task runs in an isolated, reproducible environment with real tools like a browser, bash, and file system. It supports four different agent harnesses—OpenClaw, Claude Code, Codex CLI, and Hermes Agent—to separate model capability from the surrounding scaffolding.
Who it’s for
It is designed for AI researchers and developers building autonomous agents who need to measure how models perform in practical, long-horizon workflows rather than simple instruction-following tests.
Highlights
- Real-world environments: Uses live tools (browser, email, calendar) instead of mocks.
- Diverse task suite: Includes multimodal tasks like video highlight clipping and complex coding tasks like debugging undocumented codebases.
- Multi-harness evaluation: Allows comparison of the same model across different agent scaffolds.
- Robustness testing: Includes safety alignment tasks like detecting leaked API keys and resisting prompt injection.
- Reproducible results: Uses isolated Docker containers and prevents data leakage by injecting grading scripts only after execution.