Z | Enterprise AI Agents @Z_AI_Lab

Anthropicの「AI-Native SDLC Playbook」がかなり参考になった。 ポイントは、AIにコードを書かせることではなく、AI Agentを前提にSDLC全体を再設計すること。 ・intent → spec → plan → code/tests → PR → deploy → incident のArtifact駆動 ・Skillsで組織知を実行可能にする ・Hooksで権限・HITL・Guardrailを決定論的に制御 ・コードだけでなくCoding Agent自体もContinuous Evalsで回帰テスト ・Parallel SessionとSubagentを使い分ける ・Incidentを次のintent / evalにつなげて閉ループ化 特に印象的だったのは、**「Agentが速くなるほど、人間のボトルネックは実装から判断・レビュー・承認へ移る」**という点。 The loop keeps running. Human judgement stays above it. 引用元：claude.com/blog/the-ai-na…

This BBC piece is worth reading. The slow adoption of AI in Japanese companies may not be because they don’t want AI, but because they don’t trust it enough to hand over execution. Writing, summarization, and information retrieval are fine. But once AI enters core operations, customer-facing workflows, or decision-making, tolerance for mistakes, unclear accountability, and reputational risk drops close to zero. That’s why I increasingly think the real challenge for Enterprise AI Agents is not just making agents smarter. It’s building Controlled Autonomy: Permission / HITL / Evaluation / Replay / Audit The goal is to gradually expand what AI can safely take responsibility for, without forcing companies to give up control. The gap between an AI Assistant and a Trusted AI Worker may be one of the most important layers to solve for enterprise AI in Japan. bbc.com/news/articles/…

Yesterday, I wrote about why enterprise AI agents need bounded autonomy, not unlimited autonomy. After spending more time studying Palantir today, I think I have a more concrete way to frame the problem: Enterprises do not need to make the LLM itself deterministic. Instead, determinism should live around the agent. The LLM can remain responsible for: - understanding intent - planning - selecting tools - generating parameters But before anything actually changes a business system, the surrounding execution layer should enforce: business rules, permissions, validation, HITL, auditability, and execution controls. For example, consider a simple rule: A closed contract cannot be modified. One approach is to put that rule into a knowledge base, retrieve it through RAG, and hope the LLM remembers to follow it every time. That is still probabilistic. A more reliable design is: The agent may attempt to call updateContract, but the execution layer deterministically checks: contract.status != CLOSED If the condition fails, the action is rejected. This is making me rethink what matters most in enterprise agent architecture. Maybe the key is not making the agent itself increasingly intelligent. Maybe it is building a reliable enterprise execution layer around it: Let the agent reason freely, but only allow it to change the real world within explicitly defined business rules, permissions, and action boundaries. Perhaps the core of controlled autonomy is not inside the agent at all. It is in the system around it.

The biggest challenge in Enterprise AI isn’t intelligence. It’s trust. While building an internal work-style AI agent on top of our company’s data platform, I realized something interesting. The problem isn’t just that LLM outputs are probabilistic. The workflow itself can also be probabilistic. Traditional workflow tools (such as Dify) follow a predefined execution path: → Input → Step A → Step B → Output The path is constrained, predictable, and relatively easy to test and audit. A Deep Agent works differently. Given a goal, it decides how to solve the problem. It plans, chooses tools, evaluates intermediate results, revises its approach, retries when necessary, and continues until it believes the task is complete. That means uncertainty exists at two levels: • What will the model output? • How will the agent execute the task? For personal productivity, this flexibility is a huge advantage. For enterprise business workflows, however, organizations also need predictable execution, clear boundaries, auditability, and accountability. Maybe the future of Enterprise AI isn’t fully autonomous or fully deterministic. It’s bounded autonomy. Curious how other teams are balancing agent flexibility with enterprise reliability.

AIエージェントにコード全部任せる（out of the loop）vs 全行レビューする（in the loop）、どっちも現場でやると微妙なんよな。 実際やってて一番しっくり来るのは "on the loop" — エージェントの動き方自体を設計・改善し続けるポジション。 プロンプトやワークフローやガードレールを磨き続けることで、アウトプットの質が底上げされていく感覚がある。 Martin Fowlerのサイトでまさにこれを言語化してる記事が出てた。Harness Engineeringって呼んでる。 martinfowler.com/articles/explo… #AIエージェント #LLM #ソフトウェア開発

金融系企業でLangGraph/CrewAIを使ったAIエージェントを実際に導入した経験があります。日本企業のAI活用について情報交換したい方、ぜひつながりましょう🤝 #AIエージェント #LLM #生成AI

ByteDance発のOSS「OpenViking」について記事を書きました。 従来のRAGが抱えるチャンク断片化・Token浪費の課題に対して、L0/L1/L2の3層構造で段階的にコンテキストをロードするアプローチが面白いです。 LangGraphでの応用パターンも紹介しています。 zenn.dev/zhu_tech/artic… #OpenViking #RAG #AIAgent #LangGraph #LLM

業務SaaSでText2SQLのPoCをやって見えた、本番化への4つの壁をまとめました。 LangGraphのHITLやベクトルDBでのスキーマ管理など、実践知見を共有しています。 zenn.dev/zhu_tech/artic… #Text2SQL #LLM #LangGraph #AI