OpenClaw把runtime、plugins、system calling和local-first memory放进同一个框架里，既回应了AI Agent工程化落地的需求，也开始触及个性化AI的方向。本文讨论它已经做对了什么、为什么这仍然只是早期实验，以及AI OS与personal memory未来可能如何分工。

从OpenAI o1与相关解读出发，这篇文章整理了对推理模型的理解，包括隐式CoT、类MCTS搜索、自举数据、Critic模型，以及推理能力与Agent能力之间的权衡。

这篇文章围绕 Plan Search 展开，讨论为什么在生成答案之前先生成高层规划，有机会提升pass@k表现，并把这个思路和CoT多样性、代码生成以及Agent工作流联系起来。

This page uses a custom Hugo layout to expose a lightweight public backlog, collecting follow-up tasks and unfinished ideas related to the blog and adjacent projects.

This post sketches an evaluation plan for a CoT-based knowledge QA agent, covering hallucination control, answer relevancy, faithfulness, dataset design, and task-specific quality metrics.

This post outlines a simple view of prompt engineering for LLM CoT, arguing that concise instructions, divide-and-conquer task design, and better workflow decomposition can improve reasoning quality.