[AINews] FrontierCode: Benchmarking for Code Quality over Slop
English summary
This newsletter highlights FrontierCode, a new benchmark from Cognition that evaluates code mergeability rather than just unit test passing, with top models scoring only 13% on the hardest subset. It covers the rise of 'loops' as an agent control metaphor, improvements in agent ergonomics, and new model releases like Kimi Code and Gemma 4. The article also discusses shifts in evaluation methodology toward real-world telemetry and the ongoing race in consumer AI platforms. Additionally, it notes research directions in continual learning and optimization.
Chinese summary
本期通讯重点介绍了 Cognition 推出的新基准 FrontierCode，该基准评估代码的可合并性而非仅仅单元测试通过率，最佳模型在最难子集上仅得分 13%。文章讨论了“循环”作为智能体控制隐喻的兴起、智能体人体工程学的改进，以及 Kimi Code 和 Gemma 4 等新模型的发布。还探讨了评估方法向真实世界遥测的转变以及消费级 AI 平台的持续竞争。此外，还提到了持续学习和优化方面的研究方向。
Key points
FrontierCode benchmark focuses on code mergeability across dimensions like regression safety, cleanliness, and maintainability, with top models scoring ~13% on hardest tasks.
FrontierCode 基准关注代码的可合并性，涵盖回归安全性、整洁性和可维护性等维度，顶尖模型在最难任务上得分约 13%。
Agent 'loops' and state machines are becoming the dominant paradigm, but practitioners caution that human checkpoints remain essential.
智能体的“循环”和状态机正在成为主导范式，但实践者警告说人工检查点仍然至关重要。
Kimi Code and Kimi Work were launched as open-source coding and desktop agent products, supporting up to 300 local sub-agents.
Kimi Code 和 Kimi Work 作为开源编码和桌面智能体产品发布，支持多达 300 个本地子智能体。
Agent Arena from Arena uses real-world sessions for leaderboard evaluation, moving beyond synthetic benchmarks.
Arena 推出的 Agent Arena 利用真实会话进行排行榜评估，超越了合成基准。
Open-source RL environment protocol OpenEnv transferred to a consortium including Hugging Face and NVIDIA.
开源 RL 环境协议 OpenEnv 已转移至包括 Hugging Face 和 NVIDIA 在内的联合体。