Tencent WorkBuddy Bench: A Multi-Domain Coding-Agent Benchmark with Contamination-Resistant Task Construction

2026-07-23 · Source: Computation and Language · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Software Development & Engineering, Robotics & Autonomous Systems · Depth: Expert, quick

Summary

Tencent WorkBuddy Bench is a new multi-domain evaluation suite for coding agents, featuring a contamination-resistant task construction methodology. It utilizes a unified framework to generate distribution-informed tasks across four key work domains: Code, Web, Office, and Security. Tasks are reverse-engineered from real commits, pull requests, or business scenarios and rewritten as colloquial, role-played requests, preventing web-search discoverability of original issues. The dataset, including task directories and evaluation harnesses, is openly released, with contamination resistance maintained through its unique construction and versioning. The suite includes four distinct subsets—repository-level engineering, front-end development, office/business workflows, and red-/blue-team security—each with specific verification styles. Tasks run under a uniform protocol on CodeBuddy Code and Claude Code agent harnesses, ensuring reproducibility and auditability. A cross-model leaderboard is reported, though subset scores are not directly comparable. The benchmark was published on 2026-07-23.

Key takeaway

For AI Engineers evaluating coding agents, the Tencent WorkBuddy Bench offers a robust, contamination-resistant resource. You should consider integrating this benchmark to assess your agents' performance across diverse real-world scenarios, including Code, Web, Office, and Security domains. Its open release and uniform protocol ensure reproducibility and auditability, allowing you to confidently compare models and identify specific areas for improvement in agent capabilities.

Key insights

The Tencent WorkBuddy Bench offers a contamination-resistant, multi-domain benchmark for coding agents, ensuring robust and reproducible evaluation.

Principles

• Task construction should prevent data contamination.
• Openly releasing benchmarks enhances auditability.
• Multi-domain evaluation reveals agent versatility.

Method

Tasks are reverse-engineered from real commits/PRs/scenarios, then rewritten as colloquial, role-played requests to avoid web-search discoverability. This ensures contamination resistance.

In practice

• Evaluate coding agents across diverse domains.
• Use reverse-engineered tasks for robust testing.
• Inspect benchmark content for auditability.

Topics

• Coding Agents
• Benchmark Datasets
• Contamination Resistance
• Multi-Domain Evaluation
• Tencent WorkBuddy Bench
• Software Engineering

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Engineer

Related on AIssential

• Tencent WorkBuddy Bench: A Multi-Domain Coding-Agent Benchmark with Contamination-Resistant Task Construction — Tencent WorkBuddy Bench offers a contamination-resistant, multi-domain benchmark for coding agents, informed by real-world usage and openly auditable.
• Beyond SWE-Bench Pro - Where do Agents go from Here? — New benchmarks are crucial for advancing AI coding agents beyond simple issue resolution to complex system understanding and improvement.
• TencentCloud / TencentDB-Agent-Memory — Layered and symbolic memory architectures significantly boost LLM agent efficiency and task success.
• StructureClaw: Traceable LLM Agents and an Executable Benchmark for Structural Engineering Workflows — Artifact-centered evaluation reveals workflow failures LLM agents miss in structural engineering.
• ICAE-Bench: Evaluating Coding Agents as Interactive Project Builders — ICAE-Bench evaluates coding agents on interactive project building, moving beyond static, fully specified tasks.
• Asuka-Bench: Benchmarking Code Agents on Underspecified User Intent and Multi-Round Refinement — Real-world code generation demands iterative refinement from underspecified user intent, a capability Asuka-Bench effectively measures.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computation and Language.