Agentic AI

用Claude Code替代API调用做翻译任务：利用agentic loop实现自我纠错，用evaluation-first定义验收标准，从过程确定性转向结果确定性获得新的安全感。

Why handing translation to Claude Code works better than calling APIs directly - leveraging the agentic loop, evaluation-first mindset, and the ecosystem's runtime layer to achieve outcome certainty over process certainty.

AI"偷懒"的本质是LLM输出长度限制导致的注意力分散。Wide Research通过多轻量模型并行处理子任务、主LLM汇总的方式解决，分享为Codex构建该能力的设计思路。

Why AI slacks off on large tasks: LLM output length limitations cause attention drift. Wide Research solves this by parallelizing with lightweight models, then aggregating results with a primary LLM.

分析OpenAI Apps SDK通过_meta域绕过context window的做法如何违背MCP设计哲学，以及协议分裂成不同dialect的潜在危机。

Analyzing how OpenAI's Apps SDK extension with _meta field violates MCP's design philosophy, creating dialects that may fragment the standard like SQL or CSS.

在真实编程和调研任务中评测Kimi K2的Agentic能力：执行韧性出色，适合作为信息采集前端；但工具调用稳定性和生态适配仍有提升空间。

Testing Kimi K2 as an agentic model in real coding and research tasks: excellent execution resilience makes it ideal for information gathering, but tool integration friction needs improvement for production use.

AI帮忙写的文章空中楼阁，不是AI写得不好，是我没喂够context。作为Manager要审核输出，作为Enabler要提供长期积累。判断界限才是协作的开始。

An AI-assisted article failed not because AI wrote poorly, but because I didn't feed it enough context. As Manager, verify output; as Enabler, provide your accumulated materials. Defining boundaries is where collaboration begins.