AIエージェント 人類と協働する機械 amazon.co.jp

It was a long read, but once Chapter 11 revealed the relationship between the SECI model and AI agents, things really heated up.

The SECIモデル (SECI model) is a theoretical framework proposed by Ikujiro Nonaka and Hirotaka Takeuchi that explains the process of organizational knowledge creation through the mutual conversion of tacit knowledge and explicit knowledge.

The last book I’d read by Hiroki was “エンジニアリング組織論への招待 ” (An Invitation to Engineering Organization Theory), which is also where I first learned about the SECI model. In that earlier work, he defined engineering as “the practice of reducing uncertainty,” and in this book he discusses collaboration with AI agents — what lies beyond that.

This book introduces a practical example of an AI agent called “Shirase-kun.” It handles tasks like automatically recording and summarizing meeting minutes — essentially an AI agent taking on the process of converting organizational tacit knowledge into explicit knowledge. The perspective that AI supports the “externalization” phase of the SECI model was particularly striking.

Right now, vibe coding through chat-style interactions with tools like Claude Code and Codex is becoming widespread, but I think the future will evolve toward having multiple AI agents collaborate and work 24/7.

Perhaps humans’ main job will become building AI agents. Not just software engineering roles, but entire organizations will be redefined — I truly felt that times are about to change.

The appendix “How This Book Was Made” also contains some amazing details. Definitely worth a read.

Highlights and Notes

Below are my personal notes.

しかし実際のところ、ソフトウェアの本質は文書管理にあります。複数の人が協働して同じ文章（コード）を管理し、版（バージョン）管理システムを通じて変更を追跡し、承認プロセスを経て統合していく様子は、まさに立法府や行政機関における文書管理プロセスと酷似しています。言い換えれば、ソフトウェアは 動作する書物 なのです。

(English translation) In reality, the essence of software is document management. Multiple people collaborating to manage the same text (code), tracking changes through version control systems, and integrating through approval processes — this closely resembles the document management processes of legislatures and administrative agencies. In other words, software is a living document.

An interesting interpretation.

人類が火を利用するようになったことは、私たちの身体機能の「外部化」の始まりを象徴しています。

(English translation) Humanity’s adoption of fire symbolizes the beginning of the “externalization” of our bodily functions.

Also interesting. And of course, AI is externalization too.

ベテラン社員の頭の中にある「こういう場合はこう判断する」という経験則を、if文やアルゴリズムとして表現する。これこそが、現代のソフトウェア開発の本質なのです。 この関係性を最も端的に表現したのが、1968年にメルヴィン・コンウェイが提唱した「コンウェイの法則」です。

(English translation) Expressing the rules of thumb in veteran employees’ heads — “in this situation, make this judgment” — as if statements and algorithms. This is the essence of modern software development. The relationship was most succinctly expressed by “Conway’s Law,” proposed by Melvin Conway in 1968.

So that’s how it connects…!

記録（SoR）、顧客接点（SoE）、洞察（SoI）
AIエージェントの登場によって可能になった第4の段階である知識創造（SoK：System of Knowledge Generation）という新しい概念を提示します。

(English translation) System of Record (SoR), System of Engagement (SoE), System of Insight (SoI) — and introducing a new concept: Knowledge Generation (SoK: System of Knowledge Generation), a fourth stage made possible by the emergence of AI agents.

「しらせ君」の実践例は、SECIモデルの理論的な枠組みが実際の業務において具体的な価値を生み出すことを示しています。

(English translation) The practical example of “Shirase-kun” demonstrates that the theoretical framework of the SECI model can produce tangible value in actual business operations.

I never expected the SECI model and AI to connect like this. And through n8n no less.

ジョブクラフティングとは、働く人一人ひとりが主体的に仕事や職場の人間関係に変化を加えることで、与えられた職務から自らの仕事の経験を創り上げていくことです。
ジョブクラフティングは、3つの要素から成り立っています（図14.2）。

(English translation) Job crafting is the practice where each individual worker proactively introduces changes to their work and workplace relationships, thereby creating their own work experience from an assigned role. Job crafting consists of three elements (Figure 14.2).

The Ministry of Health, Labour and Welfare has published “2019 White Paper on the Labour Economy — Issues Surrounding ‘Ways of Working’ Under Labour Shortages ,” and “Section 3: Challenges Toward Realizing an Environment Where People Can Work with a Sense of Purpose ” contains a column on job crafting.

The following is quoted from there:

• Task crafting
  • Ingenuity in how work is done; changing the volume or scope of work so that its content becomes more fulfilling
• Relational crafting
  • Ingenuity in engaging with people around you; adjusting how you interact with those involved in your work to receive support and positive feedback, thereby increasing job satisfaction
• Cognitive crafting
  • Ingenuity in how you perceive and think about work; reframing the purpose and meaning of work, or connecting it to your own interests, so you can work proactively with a sense of fulfillment

Job crafting was a new concept for me, but I think I’ve been doing it fairly consciously in my own work.

AIエージェントが高度化し、より多くのタスクを代行できるようになればなるほど、人間のエージェンシー（主体性）はより重要になってきます。なぜなら、AIが「何をするか（What）」を実行できるようになった今、人間に求められるのは「なぜそれをするのか（Why）」「どのような価値を生み出すのか（Value）」を定義することだからです。

(English translation) The more sophisticated AI agents become and the more tasks they can handle, the more important human agency becomes. Now that AI can execute “What to do,” what’s required of humans is to define “Why do it” and “What value does it create.”

“Why” is becoming increasingly important. My awareness of “Value” has been somewhat lacking, so I’ll make a point of improving that.

ソフトウェア開発における真のボトルネックは、もっと根源的な場所にあったのです。 それは「暗黙知」です。

(English translation) The true bottleneck in software development was in a more fundamental place. It’s “tacit knowledge.”

I completely agree. With tools like Claude Code, you can quickly understand what code does, but the tacit knowledge lurking behind it is often something only the original implementer can truly grasp.

多くの組織でAI活用が「タスクの効率化」で止まってしまうのは、この暗黙知を拾い上げ、問いを立てるプロセスが不足しているからではないでしょうか。

(English translation) Perhaps the reason AI adoption stalls at “task efficiency” in many organizations is the lack of a process for surfacing this tacit knowledge and formulating the right questions.

I think so too.

Claude CodeによるAI駆動開発入門 amazon.co.jp

My vibe coding journey started with Cursor, but since around May of last year, I’ve been almost exclusively using Claude Code. Even writing this blog, I let Claude Code take it from zero to roughly 60% done.

I’ve been keeping up with Claude Code by reading and trying out information that flows through Twitter and blog posts, in a mix of relaxation and a sense of urgency.

Much of the content was already familiar, but it helped me organize my fragmented knowledge and validate many of my assumptions.

The book walks you through building an app step by step, all the way to deploying it on Google Cloud. I found it to be beginner-friendly for those new to vibe coding.

Since I was already a fairly experienced user, I didn’t follow along hands-on, but I think it would make a great tutorial for first-timers.

This is an incredibly fast-moving field, so if you’re interested, I’d recommend buying it and trying it out right away. This is one book you definitely shouldn’t let sit on the shelf lol

Highlights and Notes

Below are my personal notes.

/compact APIの仕様についてフォーカスしてください

(English translation) /compact Focus on API specifications

I may never need it, but I learned that you can pass arguments to /compact.

It seems that slash commands in general allow you to pass arguments both explicitly and loosely.

AnthropicはSOC2 Type2レポート、ISO 27001証明書などを取得しており、詳細はTrust Center 注5.1 で確認できます。

(English translation) Anthropic has obtained SOC2 Type 2 reports, ISO 27001 certifications, and more. Details can be found at the Trust Center Note 5.1 .

I didn’t know that. This could be useful when considering adoption at a company.

touch ~/.config/tmux/tmux.conf

Not directly related to Claude Code, but I discovered that tmux now supports XDG_CONFIG_HOME.

After checking, it appears this has been the case since v3.2a .

claude -p ${プロンプト}を使うとClaude Codeをヘッドレスモードで使うことができます。

(English translation) You can use Claude Code in headless mode with claude -p ${prompt}.

I knew Gemini CLI could do this, but I didn’t realize Claude Code could too. Interesting.

さらに自身のコードを改善するフィードバックループが社内で既に標準化されており、その有用性の高さから将来的に外部開発者向けに公開する可能性も示唆されています。少なくとも本書で紹介した/security-reviewコマンドは、当初そういった社内ツールの一環だったようです。

(English translation) Furthermore, a feedback loop for improving its own code has already been standardized internally, and given its high utility, they’ve hinted at the possibility of making it available to external developers in the future. At the very least, the /security-review command introduced in this book appears to have originally been part of such internal tools.

I wonder if these are what they were referring to:

• https://github.com/anthropics/claude-code-security-review
• https://github.com/anthropics/claude-plugins-official
• https://github.com/anthropics/skills