Evaluated coding agents have always been confined to isolated fields - fixing bugs on SWE-bench, doing front-end work on Design2Code, and production environment benchmarks are mostly closed to external audits. Tencent has now brought these four areas together, releasing a multi-domain evaluation suite called WorkBuddy Bench, with the paper published on arXiv. Its most notable feature is not the number of tasks, but that these tasks are designed from the ground up to prevent "memorizing answers."
The entire benchmark spans four work domains: code (repository-level software engineering), web (frontend artifacts), office (multi-file business processes), and security (red and blue teams). The scales are 80, 70, 50, and 60 tasks respectively, totaling 260. The key point is that none of the tasks are derived from any public question bank, but rather reverse-engineered from real code commits, pull requests, or business scenarios, then rewritten into short, colloquial, role-playing-style requests. This ensures that the task prompts cannot be found by searching online for corresponding PRs or commit clues - you can't memorize them. Since the dataset is publicly released (including task directories, environment images, evaluation tools, test cases, and reference solutions), its resistance to contamination comes from this construction method plus version control, rather than keeping the questions hidden.
The four subsets share the same task directory format, but each has its own validation method, so scores cannot be directly compared between subsets. Therefore, Tencent does not report an overall score for the suite. Code tasks are scored based on hidden test pass rates; web tasks use rule checks combined with LLM/VLM evaluation and agent evaluation, requiring delivery of artifacts along a declared path without connecting to the real-time internet; office tasks use deterministic rule checks combined with evidence-based LLM evaluation, with weights ranging from 0.70 to 0.95 favoring rules; security tasks use deterministic scoring.py runs three times and take an average, without using large models as judges. The security subset also includes five layers of anti-cheating measures - disabling literal scanning, renaming inputs, covering tampered tests, encoding dependencies, and low-weight bait, with 38 red team tasks and 22 blue team tasks. White-box audits are anchored on real-world CVEs like binutils, curl, and nginx.
The task construction follows a unified pipeline: source collection, rewriting into real requests, assembling the workspace, isolating assessment assets after rounds, and packaging into independent directories. User data is never touched throughout the process. Code tasks assign five roles (developer, algorithm engineer, product manager, QA, operations), while security tasks assign professional roles, and deliberately provide incomplete information - they don't tell you the target file, pattern, or boundary, so the agent must retrieve the context itself. Scoring assets are introduced only after the agent completes its run, and it never sees them during the process.
The evaluation runs in isolated containers, with the model and sandbox separated. The framework uses CodeBuddy Code (default) and Claude Code, with increased reasoning efforts and a context window of 200k, and WebSearch and AskUserQuestion are disabled. This is important: turning off online search is precisely to verify whether the resistance to contamination is truly effective.
The leaderboard features multiple model families (scores 0–100, thinking mode, three averages). Claude Opus4.8 dominates the majority of the top positions in code, web, office, and security, taking five first places; GLM-5.2 tops two security rankings, and GPT-5.5 takes the office cc first. The framework's sensitivity is also significant - the security subset shows the most dramatic re-ranking, with an average absolute change of 8.6 points; Claude Opus4.8 rejected 13 answers under the cc framework, while GPT-5.5 rejected only 2 under cbc. In terms of efficiency, GPT-5.5 outputs the fewest tokens but maintains a decent score, while DeepSeek-V4-Pro ran 44 rounds on code, with high token input and output, appearing more "demanding."
