Tencent releases WorkBuddy Bench, a new open benchmark for testing coding agents across multiple domains

Tencent has unveiled WorkBuddy Bench, a comprehensive evaluation framework designed to assess how well coding agents perform across a range of practical work scenarios. Rather than relying on publicly available problems that could skew results through data contamination, the benchmark constructs its tasks from actual code repositories, pull requests, and business workflows—then rephrases them as natural human requests that can't be easily found through web searches.

Benchmark design and contamination resistance#

The suite covers four distinct domains: code repository engineering, front-end web development, office and business workflows, and security testing (both red- and blue-team scenarios). Each domain presents unique challenges and uses its own evaluation methodology tailored to real-world requirements.

The key innovation lies in contamination resistance. Instead of directly copying existing public issue text, each task is reverse-engineered from an actual commit or pull request, then rewritten as a conversational, role-played instruction. This approach prevents models from simply recalling answers they've seen during training. The dataset is released openly—including task directories, environment images, test harnesses, and reference solutions—so reproducibility and transparency don't depend on secrecy but rather on how the tasks themselves are constructed and versioned.

Evaluation and results#

All tasks follow a uniform task-directory format and are executed under a consistent, reproducible protocol. The benchmark has been tested on two agent harnesses: CodeBuddy Code and Claude Code. Since each domain uses a different scoring mechanism appropriate to its nature, the suite deliberately avoids reporting a single suite-wide average; instead, it publishes separate leaderboards across multiple model families, allowing direct comparison within each work domain.

The open release of the full benchmark—including source materials, environment setup, and testing infrastructure—makes it auditable by any third party. Researchers can re-run tasks independently and inspect the underlying content, establishing confidence in how the agents are being measured.