本文作者是一位资深工程师，分享了从2025年到2026年间，他使用AI/LLM方式发生的巨大变化。核心变化在于：AI代理现在已经相当成熟，从过去偶尔且谨慎地使用，转变为几乎持续使用且仅需轻度监督。例如，他如今会将几乎所有代码修改交给AI代理来完成，并利用其进行bug调查、测试和本地环境配置。然而，他依然坚持自己撰写PR描述、技术文档和公开沟通信息，并强调人类专业判断在复杂问题排查中仍然至关重要。作者认为，当前的核心AI技能是尽可能多地将工作移交给AI代理，但又要把握好不过度的分寸。
VeilGate is a deception reverse proxy that uses scoring and tarpits to defend against autonomous AI pentesting agents, unlike traditional WAFs that reveal weaknesses through block responses.
InsForge is an open-source backend platform (Apache 2.0) for AI coding agents to deploy, operate, and debug via CLI and Skills. It includes auth, databases, hosting, cron jobs, and safety features like backend branching, telemetry, and a debug agent.
Valgrind 3.27.1 has been released as a minor point update addressing a regression introduced in version 3.27. The release fixes issues including the "--fair-sched=yes" option not working and Valgrind incorrectly unpacking sys_port results, which caused significant delays.
A staff engineer reflects on their practical use of LLMs in 2026, describing them as a helpful tool for code generation, refactoring, and documentation, while emphasizing the importance of human judgment in reviewing and adapting AI-generated output.
The author, a staff engineer, describes using LLMs primarily as a code review tool and for generating boilerplate code, while emphasizing that LLMs are not yet reliable for complex architectural decisions or writing production-ready code without careful human oversight. They find the most value in using LLMs to quickly get feedback on design documents and to automate tedious coding tasks.