<?xml version="1.0" encoding="utf-8" standalone="yes"?><rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel><title>Writing on Mai</title><link>https://o-sakurajimamai-o.github.io/blog/</link><description>Recent content in Writing on Mai</description><generator>Hugo</generator><language>en-us</language><lastBuildDate>Wed, 11 Nov 2026 17:34:08 +0800</lastBuildDate><atom:link href="https://o-sakurajimamai-o.github.io/blog/index.xml" rel="self" type="application/rss+xml"/><item><title>AgentRx: Diagnosing AI Agent Failures from Execution Trajectories</title><link>https://o-sakurajimamai-o.github.io/blog/agentrx-notes/</link><pubDate>Wed, 11 Nov 2026 17:34:08 +0800</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/agentrx-notes/</guid><description>&lt;p&gt;AgentRx proposes a new attribution framework, but the overall process still has two steps: flag failures, then identify which failure was critical.&lt;/p&gt;&#10;&lt;hr&gt;&#10;&lt;p&gt;AgentRx groups failures into nine categories ($K_1$–$K_9$): instruction or plan adherence failure, intent–plan mismatch, fabrication of new information or hallucination, invalid tool call, misinterpretation of tool output, insufficient specification of user intent, unsupported intent, guardrail triggered, and system failure.&lt;/p&gt;&#10;&lt;p&gt;AgentRx takes three inputs: (i) a toolset $\mathcal{T}$ with input and output schemas, (ii) an optional domain policy $\Pi$, and (iii) a failed trajectory $T=\langle s_{1},\dots,s_{n}\rangle$. Each step $s_k$ contains multiple logged events with fields such as agent name, tool name, step index, tool output, or natural-language conversation.&lt;/p&gt;</description></item><item><title>BLADE: Auditing Information Use for Failure Attribution</title><link>https://o-sakurajimamai-o.github.io/blog/blade-notes/</link><pubDate>Tue, 10 Nov 2026 11:50:45 +0800</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/blade-notes/</guid><description>&lt;h2 id="blade"&gt;BLADE&lt;/h2&gt;&#10;&lt;p&gt;BLADE uses Jev for attribution and divides the process into three stages. Notably, Jev makes the decisions in all three stages. The diagram below shows the BLADE framework.&lt;/p&gt;&#10;&lt;figure class="fig"&gt;&#10; &lt;img src="https://o-sakurajimamai-o.github.io/blog/blade-notes/figures/blade-framework.png" width="1904" height="732" alt="Figure 1. Overview of BLADE. (1) Jev reconstructs dependencies between events. (2) The dependencies select the handoffs to audit, and Jev checks how each agent decision uses its source against the task’s requirement spans $R$. (3) The diagnosis reads the full trajectory with the blame graph $G=(V,E,R,F)$; every event remains a candidate." loading="lazy" decoding="async"&gt;&#10; &lt;figcaption&gt;Figure 1. Overview of BLADE. (1) Jev reconstructs dependencies between events. (2) The dependencies select the handoffs to audit, and Jev checks how each agent decision uses its source against the task’s requirement spans $R$. (3) The diagnosis reads the full trajectory with the blame graph $G=(V,E,R,F)$; every event remains a candidate.&lt;/figcaption&gt;&#10;&lt;/figure&gt;&#10;&lt;h2 id="method"&gt;Method&lt;/h2&gt;&#10;&lt;p&gt;BLADE represents an event as $e_i$ and defines four kinds of event relation: depends on, repeats, contradicts, and repairs. Jev selects the relation between events, with an additional None option. Let $\mathcal{R}_{\mathrm{link}}$ denote the four relation types and let $x_{t,r}$ denote the question about relation $r$ at event $e_t$. The probability assigned to an earlier event $e_u$ is&lt;/p&gt;</description></item><item><title>DoVer: Intervention-Driven Multi-Agent Debugging</title><link>https://o-sakurajimamai-o.github.io/blog/dover-notes/</link><pubDate>Fri, 06 Nov 2026 18:45:50 +0800</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/dover-notes/</guid><description>&lt;h2 id="the-problem"&gt;The Problem&lt;/h2&gt;&#10;&lt;p&gt;The evaluation approach used by the Who&amp;amp;When (WW) dataset has limitations. Its setup is:&lt;/p&gt;&#10;&lt;ol&gt;&#10;&lt;li&gt;&lt;strong&gt;Input:&lt;/strong&gt; A complete failed session log, from the user&amp;rsquo;s question through rounds of interaction among agents to an incorrect final answer or a stopping rule such as the maximum number of replanning rounds.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;Task:&lt;/strong&gt; Ask an LLM to read the log and identify the earliest, uniquely responsible agent and step that caused the task to fail.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;Decisive error step:&lt;/strong&gt; A step whose incorrect action could be replaced by a correct one so that the subsequent trajectory succeeds.&lt;/li&gt;&#10;&lt;/ol&gt;&#10;&lt;p&gt;The authors found that WW trajectories did not explicitly show step numbers, so the LLM had to track or infer each step. The evaluation criteria given to the LLM also differed from those used in expert annotations. The authors embedded a concise version of the human annotator guidance in the LLM prompt: &lt;code&gt;Find the earliest decisive step that diverted the task and whose correction could lead to success&lt;/code&gt;.&lt;/p&gt;</description></item><item><title>TrajDebug: Tracing Error Lifecycles</title><link>https://o-sakurajimamai-o.github.io/blog/trajdebug-notes/</link><pubDate>Thu, 05 Nov 2026 23:32:07 +0800</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/trajdebug-notes/</guid><description>&lt;h2 id="core-idea"&gt;Core Idea&lt;/h2&gt;&#10;&lt;p&gt;The paper&amp;rsquo;s central idea is to find the &lt;strong&gt;critical error step&lt;/strong&gt; $t^\star$ in a failed trajectory $\tau = (x_1, x_2, \dots, x_N)$. The authors identify $t^\star$ as the &lt;strong&gt;earliest error step directly and causally responsible for the final task failure&lt;/strong&gt;: if that step had been corrected, the later cascade would not have occurred and the task would have succeeded.&lt;/p&gt;&#10;&lt;p&gt;For an evaluation such as Who&amp;amp;When that directly gives an LLM a failed trajectory, long context, numerous local errors, and cascading mistakes often lead to poor performance.&lt;/p&gt;</description></item><item><title>Multi-Agent Failure Attribution Benchmark Notes</title><link>https://o-sakurajimamai-o.github.io/blog/multi-agent-attribution-benchmarks/</link><pubDate>Wed, 04 Nov 2026 21:57:38 +0800</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/multi-agent-attribution-benchmarks/</guid><description>&lt;h2 id="whowhen"&gt;Who&amp;amp;When&lt;/h2&gt;&#10;&lt;p&gt;Published in mid-2025, the paper uses relatively older models. The table shows the performance of OpenAI o1 and DeepSeek R1 on the dataset:&lt;/p&gt;&#10;&lt;figure class="fig"&gt;&#10; &lt;img src="https://o-sakurajimamai-o.github.io/blog/multi-agent-attribution-benchmarks/figures/who-when-reasoning-models.png" width="1464" height="244" alt="Table 1. Failure attribution performance of OpenAI o1, DeepSeek R1, and GPT-4o" loading="lazy" decoding="async"&gt;&#10; &lt;figcaption&gt;Table 1. Failure attribution performance of OpenAI o1, DeepSeek R1, and GPT-4o&lt;/figcaption&gt;&#10;&lt;/figure&gt;&#10;&lt;p&gt;The dataset contains algorithm-generated and hand-crafted agent systems: 127 systems in total, with human annotations. It distinguishes two levels of error: Agent-Level (Who, identifying the responsible agent) and Step-Level (When, identifying the specific step). Step-level performance is lower.&lt;/p&gt;</description></item><item><title>Multi-Agent Systems Survey Notes</title><link>https://o-sakurajimamai-o.github.io/blog/multi-agent-survey-notes/</link><pubDate>Fri, 23 Oct 2026 00:00:00 +0800</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/multi-agent-survey-notes/</guid><description>&lt;p&gt;Recent LLM capabilities have grown linearly or exponentially, sometimes exceeding even AI developers&amp;rsquo; expectations. OpenAI reported that an internal model coordinated around 10,000 concurrent agents to produce &lt;a href="https://openai.com/index/navier-stokes-solution/"&gt;a solution to the Navier–Stokes Millennium Prize problem&lt;/a&gt;. Anthropic researchers also described &lt;a href="https://www.anthropic.com/research/claude-shaped-science"&gt;genetics research conducted with Claude and domain experts&lt;/a&gt;. Google&amp;rsquo;s Gemini 4 Argon helped researchers &lt;a href="https://www.techradar.com/pro/its-over-ai-won-google-gemini-4-argon-shatters-59-year-old-nuclear-fusion-conjecture-in-just-a-few-hours"&gt;extend mathematical solutions related to magnetic confinement in fusion&lt;/a&gt; through multi-agent collaboration. This work relates to Grad&amp;rsquo;s 59-year-old conjecture.&lt;/p&gt;&#10;&lt;p&gt;As task scale and execution time increase, it becomes harder to concentrate every required capability in a single agent. Multi-agent collaboration offers a principled way to address this challenge.&lt;/p&gt;</description></item><item><title>Graph-based Agent Memory: Reading Notes</title><link>https://o-sakurajimamai-o.github.io/blog/graph-based-agent-memory/</link><pubDate>Tue, 20 Oct 2026 00:00:00 +0800</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/graph-based-agent-memory/</guid><description>&lt;style&gt;&#10;.prose mjx-container[display="true"] { max-width: 100%; overflow-x: auto; overflow-y: hidden; padding: .4em 0; }&#10;&lt;/style&gt;&#10;&lt;h2 id="the-agent-memory-lifecycle"&gt;The Agent Memory Lifecycle&lt;/h2&gt;&#10;&lt;p&gt;The lifecycle consists of four stages: memory extraction, storage, retrieval, and evolution.&lt;/p&gt;&#10;&lt;ol&gt;&#10;&lt;li&gt;Memory extraction: transforming raw, unstructured observations $o_t$ into structured memory units $m$.&lt;/li&gt;&#10;&lt;li&gt;Memory storage: indexing and organizing the extracted units within a memory structure $\mathcal{M}$.&lt;/li&gt;&#10;&lt;li&gt;Memory retrieval: retrieving relevant stored information $\mathcal{M}_{\text{rel}}\subset\mathcal{M}$ in response to a query $q$.&lt;/li&gt;&#10;&lt;li&gt;Memory evolution: refining memory through post-processing, including internal self-evolution, such as consolidation and abstraction, and external self-exploration, such as new environmental feedback.&lt;/li&gt;&#10;&lt;/ol&gt;&#10;&lt;figure class="fig"&gt;&#10; &lt;img src="https://o-sakurajimamai-o.github.io/blog/graph-based-agent-memory/figures/F001.png" width="2590" height="910" alt="Figure 1. Workflow of an AI agent system and detailed implementations of the memory system." loading="lazy" decoding="async"&gt;&#10; &lt;figcaption&gt;Figure 1. Workflow of an AI agent system and detailed implementations of the memory system.&lt;/figcaption&gt;&#10;&lt;/figure&gt;&#10;&lt;h2 id="a-multidimensional-taxonomy-of-agent-memory"&gt;A Multidimensional Taxonomy of Agent Memory&lt;/h2&gt;&#10;&lt;p&gt;Agent memory can be classified along several dimensions, including temporal scope, functional role, and representational structure.&lt;/p&gt;</description></item><item><title>Dream-RSI: Reading Notes</title><link>https://o-sakurajimamai-o.github.io/blog/dream-rsi/</link><pubDate>Fri, 16 Oct 2026 00:00:00 +0800</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/dream-rsi/</guid><description>&lt;style&gt;&#10;.prose mjx-container[display="true"] { max-width: 100%; overflow-x: auto; overflow-y: hidden; padding: .4em 0; }&#10;&lt;/style&gt;&#10;&lt;h2 id="pain-points"&gt;Pain Points&lt;/h2&gt;&#10;&lt;ol&gt;&#10;&lt;li&gt;Evaluating a candidate program usually only requires running it and checking the result. Evaluating an exploration policy requires letting it schedule many successive attempts, then observing the best solution found and the resources consumed. A direction with a low initial score may produce a breakthrough after several repairs; an early leader may soon stagnate.&lt;/li&gt;&#10;&lt;li&gt;Feedback on policy quality is more delayed than feedback on individual programs. Rerunning the entire online discovery process after every controller change repeatedly incurs costs for model calls, code execution, and evaluation. Meanwhile, the space of exploration policies is large. Branch selection, stopping rules, parallel batches, and failure handling can all be combined in different ways. These are the two bottlenecks identified in the introduction: &lt;strong&gt;expensive, delayed feedback and many candidate meta-policies&lt;/strong&gt;.&lt;/li&gt;&#10;&lt;li&gt;A fixed exploration policy is simple and stable, but cannot change its search habits across rounds based on accumulated discovery experience. Dream-RSI aims to reduce the cost of comparing exploration policies, rather than directly designing a stronger fixed heuristic for a particular domain.&lt;/li&gt;&#10;&lt;/ol&gt;&#10;&lt;h2 id="how-dream-rsi-works"&gt;How Dream-RSI Works&lt;/h2&gt;&#10;&lt;h3 id="0-dream-rsi"&gt;0. Dream-RSI&lt;/h3&gt;&#10;&lt;p&gt;Dream-RSI consists of online exploration and offline planning, alternating through the following cycle:&lt;/p&gt;</description></item><item><title>PPO to GRPO</title><link>https://o-sakurajimamai-o.github.io/blog/ppo-to-grpo/</link><pubDate>Sun, 11 Oct 2026 00:00:00 +0000</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/ppo-to-grpo/</guid><description>&lt;h2 id="the-ppo-algorithm"&gt;The PPO Algorithm&lt;/h2&gt;&#10;&lt;p&gt;Let&amp;rsquo;s start by defining a few concepts as the foundation for everything that follows:&lt;/p&gt;&#10;&lt;ol&gt;&#10;&lt;li&gt;&lt;strong&gt;Action&lt;/strong&gt;: the set of selectable actions, i.e., the moves the model can take next.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;Policy&lt;/strong&gt;: a policy function that takes a state as input and outputs a probability distribution over actions, usually denoted by $\pi$, e.g. $\pi(action1|S_t)=0.78\dots$&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;Trajectory&lt;/strong&gt;: denoted by $\tau$, a sequence of states and actions, where $S_{t+1}=f(S_t,a_t)$ is deterministic, or $S_{t+1}=f(*|S_t,a_t)$ is stochastic.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;Return&lt;/strong&gt;: the cumulative sum of rewards from the current time step until the end.&lt;/li&gt;&#10;&lt;/ol&gt;&#10;&lt;p&gt;Expectation:&lt;/p&gt;</description></item><item><title>KL Divergence</title><link>https://o-sakurajimamai-o.github.io/blog/kl-divergence/</link><pubDate>Fri, 09 Oct 2026 00:00:00 +0800</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/kl-divergence/</guid><description>&lt;style&gt;&#10;.prose mjx-container[display="true"] { max-width: 100%; overflow-x: auto; overflow-y: hidden; padding: .4em 0; }&#10;.prose .kl-table { overflow-x: auto; }&#10;.prose .kl-table table { width: 100%; min-width: 24rem; }&#10;&lt;/style&gt;&#10;&lt;p&gt;After SFT, a model has been trained at the token level, making its token probability distribution favor responses that follow the intended format. RL can further improve how it answers and the accuracy of its responses. During RL training, we do not want the model to stray far from the response format established by SFT, so we use KL divergence as a constraint to limit how far the model moves from the post-SFT model while trying to further improve accuracy.&lt;/p&gt;</description></item><item><title>LoRA</title><link>https://o-sakurajimamai-o.github.io/blog/lora/</link><pubDate>Thu, 08 Oct 2026 00:00:00 +0800</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/lora/</guid><description>&lt;style&gt;&#10;.prose mjx-container[display="true"] { max-width: 100%; overflow-x: auto; overflow-y: hidden; padding: .4em 0; }&#10;.prose .lora-table { overflow-x: auto; }&#10;.prose .lora-table table { width: 100%; min-width: 24rem; }&#10;&lt;/style&gt;&#10;&lt;p&gt;Fine-tuning an entire model directly requires substantial computing resources and GPU memory. For Qwen3-0.6B (0.6 billion parameters), memory requirements cannot be determined from FP16 or FP32 alone: they also depend on the optimizer, sequence length, batch size, and techniques such as gradient checkpointing or offloading. For larger models, such as 7B or 13B, full fine-tuning on a single consumer GPU generally becomes more difficult.&lt;/p&gt;</description></item><item><title>MCP, A2A, ANP</title><link>https://o-sakurajimamai-o.github.io/blog/mcp-a2a-anp/</link><pubDate>Mon, 05 Oct 2026 00:00:00 +0000</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/mcp-a2a-anp/</guid><description>&lt;h2 id="mcp-model-context-protocol"&gt;MCP (Model Context Protocol)&lt;/h2&gt;&#10;&lt;p&gt;MCP has three layers: the upstream, the MCP protocol, and the downstream.&lt;/p&gt;&#10;&lt;p&gt;The downstream is called the MCP Server and is provided by developers or vendors. It implements the specific business logic and registers and exposes three core types of assets through the MCP specification:&lt;/p&gt;&#10;&lt;ul&gt;&#10;&lt;li&gt;&lt;strong&gt;Tools&lt;/strong&gt;: functions that the model can call, such as Git operations, SQL queries, and calls to internal microservices.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;Resources&lt;/strong&gt;: data context that the model can attach in read-only form, such as local files, table schemas, and log streams.&lt;/li&gt;&#10;&lt;li&gt;&lt;strong&gt;Prompts&lt;/strong&gt;: standardized prompt templates and workflow presets for interactions.&lt;/li&gt;&#10;&lt;/ul&gt;&#10;&lt;p&gt;The upstream is called the MCP Client or host environment: commonly used agent engines such as Codex, Claude Code, and Pi. As clients, they manage lifecycles, protocol translation, and permissions:&lt;/p&gt;</description></item><item><title>Mini-SWE-Agent Source Code Analysis</title><link>https://o-sakurajimamai-o.github.io/blog/mini-swe-agent/</link><pubDate>Sun, 20 Sep 2026 00:00:00 +0000</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/mini-swe-agent/</guid><description>&lt;p&gt;After setting up the environment in &lt;code&gt;.config/mini-swe-agent&lt;/code&gt;, we run:&lt;/p&gt;&#10;&lt;div class="code-box"&#10; data-code-box data-preview="7"&gt;&#10; &lt;div class="code-bar"&gt;&#10; &lt;span class="code-lang"&gt;bash&lt;/span&gt;&#10; &lt;span class="code-lines"&gt;5 lines&lt;/span&gt;&#10; &lt;span class="code-bar__gap"&gt;&lt;/span&gt;&#10; &lt;button type="button" class="code-btn code-copy"&#10; title="Copy" aria-label="Copy"&gt;&#10; &lt;svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"&#10; stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"&gt;&#10; &lt;rect x="9" y="9" width="13" height="13" rx="2"/&gt;&#10; &lt;path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1"/&gt;&#10; &lt;/svg&gt;&#10; &lt;/button&gt;&#10; &#10; &lt;/div&gt;&#10;&#10; &lt;div class="code-body"&gt;&#10; &lt;div class="highlight"&gt;&lt;pre tabindex="0" class="chroma"&gt;&lt;code class="language-bash" data-lang="bash"&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;python -m minisweagent.run.mini &lt;span class="se"&gt;\&#10;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt; -c mini.yaml &lt;span class="se"&gt;\&#10;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt; -t &lt;span class="s2"&gt;&amp;#34;Create a test_agent.txt file and write Hello Agent inside it&amp;#34;&lt;/span&gt; &lt;span class="se"&gt;\&#10;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt; --model &lt;span class="s2"&gt;&amp;#34;deepseek/deepseek-chat&amp;#34;&lt;/span&gt; &lt;span class="se"&gt;\&#10;&lt;/span&gt;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt; --agent-class minisweagent.agents.default.DefaultAgent&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&#10; &#10; &lt;/div&gt;&#10;&#10; &#10;&lt;/div&gt;&#10;&lt;p&gt;This produces the latest execution trajectory in &lt;code&gt;.config/mini-swe-agent&lt;/code&gt;. We can interpret the source code based on this trajectory.&#10;The framework of the entire Repo:&#10;&lt;code&gt;run/mini.py:99-101&lt;/code&gt;:&lt;/p&gt;</description></item><item><title>AI Agent Learning Resources &amp; Roadmap</title><link>https://o-sakurajimamai-o.github.io/blog/ai-agent-resources-for-now/</link><pubDate>Wed, 16 Sep 2026 00:00:00 +0000</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/ai-agent-resources-for-now/</guid><description>&lt;ol&gt;&#10;&lt;li&gt;A Markdown guide consolidating quality community posts, official blogs, research papers, open-source projects, and hands-on engineering practices into an actionable AI Agent learning checklist: &lt;a href="https://github.com/datawhalechina/Agent-Learning-Hub"&gt;Agent-Learning-Hub&lt;/a&gt;&lt;/li&gt;&#10;&lt;li&gt;Anthropic&amp;rsquo;s engineering blog—a must-read is &lt;strong&gt;&amp;ldquo;Building Effective Agents&amp;rdquo;&lt;/strong&gt;: &lt;a href="https://www.anthropic.com/engineering"&gt;Anthropic Engineering&lt;/a&gt;&lt;/li&gt;&#10;&lt;li&gt;A highly-starred, open-source systematic tutorial on building and implementing agent systems from scratch: &lt;a href="https://github.com/datawhalechina/Hello-Agents"&gt;Hello-Agents&lt;/a&gt;&lt;/li&gt;&#10;&lt;li&gt;Featured post - &lt;a href="https://www.nowcoder.com/discuss/847995166416703488?sourceSSR=users"&gt;Comprehensive Roadmap for AI Application Development&lt;/a&gt;&lt;/li&gt;&#10;&lt;li&gt;Harness-pattern agents: highly recommended to dive directly into the source code of Claude Code and Pi: &lt;a href="https://github.com/shareAI-lab/learn-claude-code"&gt;Learn Claude Code&lt;/a&gt;; an insightful deep dive into the Pi agent: &lt;a href="https://mariozechner.at/posts/2025-11-30-pi-coding-agent/"&gt;pi-coding-agent&lt;/a&gt;&lt;/li&gt;&#10;&lt;li&gt;The LangChain ecosystem: after understanding the core &lt;a href="https://github.com/langchain-ai/langchain"&gt;Langchain&lt;/a&gt; library, try exploring and reproducing &lt;a href="https://github.com/langchain-ai/langgraph"&gt;LangGraph&lt;/a&gt; and &lt;a href="https://github.com/langchain-ai/deepagents"&gt;DeepAgents&lt;/a&gt;; leverage &lt;a href="https://www.langchain.com/langsmith/observability"&gt;LangSmith&lt;/a&gt; for trace monitoring, trajectory inspection, and agent evaluation.&lt;/li&gt;&#10;&lt;li&gt;Agent evaluation: an LLM-as-a-Judge framework tailored for LLMs and AI Agents; currently, &lt;a href="https://github.com/agentscope-ai/OpenJudge"&gt;OpenJudge&lt;/a&gt; stands out as a remarkably comprehensive open-source project.&lt;/li&gt;&#10;&lt;/ol&gt;</description></item><item><title>408 Exam Notes: Data Structures, Computer Organization, OS, and Networks</title><link>https://o-sakurajimamai-o.github.io/blog/kaoyan-408/</link><pubDate>Wed, 09 Sep 2026 00:00:00 +0000</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/kaoyan-408/</guid><description>&lt;p&gt;These notes organize concepts, examples, derivations, and revision reminders by topic, with the original figures. Apply each result under its stated assumptions and the model and data specified by the exercise.&lt;/p&gt;&#10;&lt;!-- source-content:cs408:start --&gt;&#10;&lt;!-- source: cs408:L1-L33 --&gt;&#10;&lt;ul&gt;&#10;&lt;li&gt;Data Structures&#10;&lt;ol&gt;&#10;&lt;li&gt;Linked lists&lt;/li&gt;&#10;&lt;li&gt;Disjoint-set union (code, optimization operations, and visualization of root merging)&lt;/li&gt;&#10;&lt;li&gt;Implementations of Prim&amp;rsquo;s and Dijkstra&amp;rsquo;s algorithms without priority queues&lt;/li&gt;&#10;&lt;li&gt;Simulation of orthogonal lists and adjacency multilists (be able to draw them)&lt;/li&gt;&#10;&lt;li&gt;Heap implementation code, and code that uses a heap to implement a priority queue&lt;/li&gt;&#10;&lt;li&gt;Red-black trees (definitions and conclusions, insertion and deletion)&lt;/li&gt;&#10;&lt;li&gt;Loser trees and replacement selection (loser tree + sequence numbers), the role of loser trees in merging, and the depth of a loser tree&lt;/li&gt;&#10;&lt;/ol&gt;&#10;&lt;/li&gt;&#10;&lt;li&gt;Computer Organization&#10;&lt;ol&gt;&#10;&lt;li&gt;Special floating-point values and flag generation&lt;/li&gt;&#10;&lt;li&gt;The principles and circuits of multiplication and division, the number of shifts in Booth multiplication and division, and rounding&lt;/li&gt;&#10;&lt;li&gt;The procedure-call process, stack frames, and their respective responsibilities&lt;/li&gt;&#10;&lt;li&gt;Multiprocessors and modern processors&lt;/li&gt;&#10;&lt;li&gt;A comprehensive summary of the single-cycle datapath in Yuan&amp;rsquo;s textbook&lt;/li&gt;&#10;&lt;/ol&gt;&#10;&lt;/li&gt;&#10;&lt;li&gt;Operating Systems&#10;&lt;ol&gt;&#10;&lt;li&gt;Interprocess communication methods&lt;/li&gt;&#10;&lt;li&gt;Simplifying resource-allocation graphs and determining the conditions for deadlock&lt;/li&gt;&#10;&lt;li&gt;Virtual machines and the detailed SPOOLing process&lt;/li&gt;&#10;&lt;li&gt;File-system mounting&lt;/li&gt;&#10;&lt;li&gt;Bitmaps and the grouped linked-list method&lt;/li&gt;&#10;&lt;li&gt;Multiprocessor scheduling&lt;/li&gt;&#10;&lt;li&gt;Wear leveling&lt;/li&gt;&#10;&lt;/ol&gt;&#10;&lt;/li&gt;&#10;&lt;li&gt;Computer Networks&#10;&lt;ol&gt;&#10;&lt;li&gt;The roles of, and differences between, the layers of OSI and TCP/IP&lt;/li&gt;&#10;&lt;li&gt;Physical-layer line coding for digital data, such as Manchester encoding; distinguish this from analog-signal modulation.&lt;/li&gt;&#10;&lt;li&gt;PPP and P2P protocols&lt;/li&gt;&#10;&lt;li&gt;Virtual circuits&lt;/li&gt;&#10;&lt;li&gt;OSPF, BGP, and DHCP messages and their specific addresses&lt;/li&gt;&#10;&lt;li&gt;The IP multicast process and IGMP&lt;/li&gt;&#10;&lt;li&gt;Hamming-code checking and error correction&lt;/li&gt;&#10;&lt;li&gt;Reservation-time calculations and NAV-value calculations in 802.11&lt;/li&gt;&#10;&lt;li&gt;Mobile IP&lt;/li&gt;&#10;&lt;/ol&gt;&#10;&lt;/li&gt;&#10;&lt;/ul&gt;&#10;&lt;!-- source: cs408:L43-L46 --&gt;&#10;&lt;h2 id="cs408-h-01"&gt;Data Structures&lt;/h2&gt;&#10;&lt;blockquote&gt;&#10;&lt;p&gt;Things to remember: the various implementation procedures for external sorting; the three types of threaded binary trees; the applicability, purposes, and drawing methods of orthogonal lists and adjacency multilists; be able to write out the complete manual calculation processes for Dijkstra&amp;rsquo;s algorithm and the critical path; insertion and deletion in balanced trees, red-black trees, and B-trees; loser-tree sequence numbers and their use in replacement-selection sorting&lt;/p&gt;</description></item><item><title>Graduate Entrance Exam Math Notes: Calculus and Linear Algebra</title><link>https://o-sakurajimamai-o.github.io/blog/kaoyan-math/</link><pubDate>Wed, 09 Sep 2026 00:00:00 +0000</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/kaoyan-math/</guid><description>&lt;p&gt;These notes organize concepts, examples, derivations, and revision reminders by topic, with the original figures. Apply each result under its stated assumptions and the model and data specified by the exercise.&lt;/p&gt;&#10;&lt;!-- source-content:calculus:start --&gt;&#10;&lt;!-- source: calculus:L1-L4 --&gt;&#10;&lt;blockquote&gt;&#10;&lt;p&gt;When working on a new problem, recognize its similarity to a previous problem. While solving it, discover gaps in your understanding of this type of problem, its tools, and its methods, or find that your understanding is not deep enough. By solving the new problem, update and refine your previous understanding at the same time.&lt;/p&gt;</description></item><item><title>Daily Codeforces Problem</title><link>https://o-sakurajimamai-o.github.io/blog/daily-question/</link><pubDate>Mon, 07 Sep 2026 00:00:00 +0000</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/daily-question/</guid><description>&lt;h2 id="array-replacement---rating-1700"&gt;Array Replacement - Rating 1700&lt;/h2&gt;&#10;&lt;p&gt;Today is Thursday, August 27, 2026. Problem link: &lt;a href="https://codeforces.com/contest/2252/problem/D"&gt;Array Replacement&lt;/a&gt;.&lt;/p&gt;&#10;&lt;p&gt;The approach to this problem is somewhat subtle, so we can only approach it from the perspective of the operation itself. This is actually a classic swap-difference problem. Consider a triplet ($a_{i-1}$, $a_i$, $a_{i+1}$). After performing one operation on it, we obtain ($a_i$, $a_{i - 1} - a_i + a_{i + 1}$, $a_{i + 1}$).&lt;/p&gt;&#10;&lt;p&gt;Let’s consider under what conditions replacing $a_i$ with $a_{i - 1} - a_i + a_{i + 1}$ is optimal. Clearly, the replacement should be made when $a_i &lt; a_{i - 1} - a_i + a_{i + 1}$. We can rearrange this expression to $a_i - a_{i + 1} &lt; a_{i - 1} - a_i$. Noting that this is very similar to the difference array, we consider defining an array $D$ such that $D_i = a_{i - 1} - a_i$.&#10;Let’s examine how operations on $a_i$ affect the array $D$:&lt;/p&gt;</description></item><item><title>NOTE：Retrieval-Augmented Generation</title><link>https://o-sakurajimamai-o.github.io/blog/noterag/</link><pubDate>Sun, 30 Aug 2026 00:00:00 +0000</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/noterag/</guid><description>&lt;h2 id="overview"&gt;Overview&lt;/h2&gt;&#10;&lt;p&gt;Paper link: &lt;a href="https://arxiv.org/abs/2005.11401"&gt;Retrieval-Augmented Generation&lt;/a&gt;&lt;/p&gt;&#10;&lt;figure class="fig"&gt;&lt;img src="https://o-sakurajimamai-o.github.io/blog/noterag/figures/overview_hu_ba8a98cb4b25f5a3.webp" width="1600" height="471"&#10; alt="Overall RAG architecture"&#10; loading="lazy" decoding="async"&gt;&lt;figcaption&gt;Overall RAG architecture&lt;/figcaption&gt;&lt;/figure&gt;&#10;&#10;&lt;p&gt;The overall process is as follows: for a specific question, the system first uses MIPS to retrieve relevant documents from the knowledge base, then &amp;ldquo;passes&amp;rdquo; them to the large model, which combines that knowledge to produce the final answer.&lt;/p&gt;&#10;&lt;p&gt;&lt;strong&gt;Questions&lt;/strong&gt;: the concrete details of the MIPS method; how the documents are passed to the large model (by augmenting the context, by modifying the query, or by making changes at the token level); what marginalization is, and why it is needed.&lt;/p&gt;</description></item><item><title>Algorithm Notes: The Grand Collection</title><link>https://o-sakurajimamai-o.github.io/blog/algorithm-collection/</link><pubDate>Sat, 29 Aug 2026 00:00:00 +0000</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/algorithm-collection/</guid><description>&lt;blockquote&gt;&#10;&lt;p&gt;This collection reorganizes the algorithm notes I wrote on cnblogs between 2023 and 2024 into seven chapters. The original reasoning, derivations, and code are preserved (including the asides I scribbled back then); places that only had code now carry explanations, and places that ran long were trimmed slightly. For the unabridged originals, follow the source-note links at the top of each section.&lt;/p&gt;&#10;&lt;/blockquote&gt;&#10;&lt;h2 id="chapter-1-basic-algorithms"&gt;Chapter 1: Basic Algorithms&lt;/h2&gt;&#10;&lt;h3 id="11-prefix-sums-and-difference-arrays"&gt;1.1 Prefix Sums and Difference Arrays&lt;/h3&gt;&#10;&lt;p&gt;Prefix sums answer &amp;ldquo;sum of any interval, many queries&amp;rdquo;: after preprocessing $s_i = s_{i-1} + a_i$, the sum over $[l, r]$ is $s_r - s_{l-1}$. The 2D version unfolds by inclusion-exclusion: $s_{i,j} = s_{i-1,j} + s_{i,j-1} + a_{i,j} - s_{i-1,j-1}$.&lt;/p&gt;</description></item><item><title>Harness Engineering with Hermes</title><link>https://o-sakurajimamai-o.github.io/blog/hermes-harness/</link><pubDate>Sat, 08 Aug 2026 00:00:00 +0000</pubDate><guid>https://o-sakurajimamai-o.github.io/blog/hermes-harness/</guid><description>&lt;h2 id="picking-a-simulated-business-scenario-and-building-the-data"&gt;Picking a Simulated Business Scenario and Building the Data&lt;/h2&gt;&#10;&lt;p&gt;I chose &lt;code&gt;an automated research assistant&lt;/code&gt;, with the data generated randomly by AI. The main structure looks like this:&lt;/p&gt;&#10;&lt;div class="code-box"&#10; data-code-box data-preview="7"&gt;&#10; &lt;div class="code-bar"&gt;&#10; &lt;span class="code-lang"&gt;text&lt;/span&gt;&#10; &lt;span class="code-lines"&gt;5 lines&lt;/span&gt;&#10; &lt;span class="code-bar__gap"&gt;&lt;/span&gt;&#10; &lt;button type="button" class="code-btn code-copy"&#10; title="Copy" aria-label="Copy"&gt;&#10; &lt;svg viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2"&#10; stroke-linecap="round" stroke-linejoin="round" aria-hidden="true"&gt;&#10; &lt;rect x="9" y="9" width="13" height="13" rx="2"/&gt;&#10; &lt;path d="M5 15H4a2 2 0 0 1-2-2V4a2 2 0 0 1 2-2h9a2 2 0 0 1 2 2v1"/&gt;&#10; &lt;/svg&gt;&#10; &lt;/button&gt;&#10; &#10; &lt;/div&gt;&#10;&#10; &lt;div class="code-body"&gt;&#10; &lt;div class="highlight"&gt;&lt;pre tabindex="0" class="chroma"&gt;&lt;code class="language-text" data-lang="text"&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;D:\Shixun\harness-research\&#10;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;├── data\ # simulated literature and experiment data (read-only, Worker must not modify)&#10;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;├── scripts\ # tool scripts (lit_read / exp_stats / report_export) ...&#10;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;├── evidence\ # evidence directory: research notes, analysis logs, review opinions (organized by task_id)&#10;&lt;/span&gt;&lt;/span&gt;&lt;span class="line"&gt;&lt;span class="cl"&gt;├── .hermes.md # project rules (discovered and injected by Hermes automatically)&lt;/span&gt;&lt;/span&gt;&lt;/code&gt;&lt;/pre&gt;&lt;/div&gt;&#10; &#10; &lt;/div&gt;&#10;&#10; &#10;&lt;/div&gt;&#10;&lt;h2 id="designing-the-workflows"&gt;Designing the Workflows&lt;/h2&gt;&#10;&lt;table&gt;&#10;&#9;&lt;thead&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;th&gt;Workflow&lt;/th&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;th&gt;Orchestration&lt;/th&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;th&gt;Why this approach&lt;/th&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;th&gt;Which layer it validates&lt;/th&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&lt;/thead&gt;&#10;&#9;&lt;tbody&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;&lt;strong&gt;1. Literature review notes&lt;/strong&gt;&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;Single Agent, direct execution&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;6 papers read sequentially, no dependencies and no parallel benefit; a single session can finish&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;&lt;strong&gt;Execution layer&lt;/strong&gt;: does the tool chain work, do Skills load, can audits leave a trail&lt;/td&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;&lt;strong&gt;2. Experiment plan design&lt;/strong&gt;&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;Parallel delegation via &lt;code&gt;delegate_task&lt;/code&gt;&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;The three subtasks — literature evidence / data audit / plan design — are independent of each other, and each needs its own context to avoid cross-contamination&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;&lt;strong&gt;Orchestration layer&lt;/strong&gt;: multi-agent division of labor, structured handoffs, merging by the main Agent&lt;/td&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&#9;&#9;&lt;tr&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;&lt;strong&gt;3. Full report review pipeline&lt;/strong&gt;&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;Kanban&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;Cross-session, ordered dependencies (Planner → Worker → Reviewer), must be able to block and recover after failures&lt;/td&gt;&#10;&#9;&#9;&#9;&#9;&#9;&lt;td&gt;&lt;strong&gt;State + control layers&lt;/strong&gt;: state persistence, rework loop, exception recovery&lt;/td&gt;&#10;&#9;&#9;&#9;&lt;/tr&gt;&#10;&#9;&lt;/tbody&gt;&#10;&lt;/table&gt;&#10;&lt;ol&gt;&#10;&lt;li&gt;Workflow 1 answers &amp;ldquo;&lt;strong&gt;can a single Agent actually get work done&lt;/strong&gt;&amp;rdquo; — the prerequisite for everything;&lt;/li&gt;&#10;&lt;li&gt;Workflow 2 answers &amp;ldquo;&lt;strong&gt;can multiple Agents divide work without interfering with each other&lt;/strong&gt;&amp;rdquo;;&lt;/li&gt;&#10;&lt;li&gt;Workflow 3 answers &amp;ldquo;&lt;strong&gt;can things recover when they go wrong&lt;/strong&gt;&amp;rdquo;.&lt;/li&gt;&#10;&lt;/ol&gt;&#10;&lt;h2 id="building-a-submittable-plugin"&gt;Building a Submittable Plugin&lt;/h2&gt;&#10;&lt;h3 id="creating-a-versioned-plugin"&gt;Creating a versioned plugin&lt;/h3&gt;&#10;&lt;p&gt;Added to the workspace:&lt;/p&gt;</description></item></channel></rss>