Series

A reusable enterprise playbook for shipping agents — from grading projects to pre-launch gates.

1. 01AI Agent Autonomy Levels: I Audited 28 'Agent' Projects — Only 5 Passed L0–L320 min read
2. 02Five Architecture Decisions That Determine Whether Your Customer-Service Agent Can Ship | Agentic AI in Practice (II)18 min read
3. 03Why a 70% Critic Beats a 95% Critic — A Fail-Closed Design Deep Dive | Agentic AI in Practice (III)22 min read
4. 04Deploy and Abandon — The Costliest Misconception in AI Agent Projects | Agentic AI in Practice (IV)16 min read
5. 05Containment Rate vs Resolution Rate: The Only Customer-Service AI Metric That Matters (How "98% CSAT" Gets Faked)16 min read
6. 06Agent Skills vs Knowledge Base: Why Stuffing SOPs Into RAG Doesn't Make an Agent Capable16 min read
7. 07Pytest-Green Doesn't Mean Ship-Ready: How to Actually Test an AI Agent (Dual-Track)16 min read
8. 08Intent Classification for Chatbots: Why Pure-Rule and Pure-LLM Both Fail (a 3-Tier Cascade)16 min read
9. 09Don't Let AI Agents Call APIs Directly — A 5-Layer Tool-Calling Stack + 25-API Contract Checklist17 min read
10. 10Corpus Drives Codebook — Why Your Intent Taxonomy Is Stuck at 60% and How It Evolves from 36 to 48 | Agentic AI in Practice (X)14 min read
11. 11LLM Fact-Checking with a Verifier Agent: 11 of 34 Facts in a 25-Page AI Plan Were Fabricated13 min read
12. 12The Org Chart Is the Real Architecture Diagram — 90% of Stalled Agent Projects Aren't a Tech Problem | Agentic AI in Practice (XII)14 min read
13. 13Self-Serve Rate ≠ Correct Rate — The Gates a Customer-Service Agent Must Clear Before Launch | Agentic AI in Practice (XIII)15 min read
14. 14Your Code Is Fixed. Production Isn't.12 min read
15. 15Your Knowledge Center Is a Search Box12 min read
16. 16One Brain, Two Front-Ends: Refunded Twice on Sale Day12 min read

Launch isn't the finish line — it's where the machine starts to turn. From how to draw eval samples, how many to label, and whether the labels are reliable, to judging launch, guarding against silent drift, and spinning up the data flywheel — six posts to make an agent that gets up, stays up, and gets sharper with use.

1. 01The 96% Turned Red the Moment We Split It by Channel — Two Fatal Failures of Eval Sampling13 min read
2. 0250 Rows at 96%, Ship It? Size Labeling by the CI Lower Bound, Not the Pretty Number12 min read
3. 03Your Labels Are Your Ceiling — One "Swap Half a Size Up" Gets Three Answers From Two Agents13 min read
4. 04A Pretty Accuracy Number Hid Dozens of Money-Moving Errors — How to Read the Eval to Ship13 min read
5. 05Your Dashboards Are Green While the Agent Quietly Gets Dumber — Post-Launch Silent Drift13 min read
6. 06Make the Agent Get Sharper With Use, Not Dumber — Spinning Up the Data Flywheel12 min read

A three-part field guide to shipping agents in retail: pick scenarios, build the KB, run ops.

1. 01Which 4 of Your 28 'Smart-X' AI Agent Projects to Start With — Retail Agentic AI Handbook (Part 1)20 min read
2. 02AI Agent Knowledge Base: 3-Layer Design + 4-Bucket Cost Estimate25 min read
3. 0380% of Failed AI Agents Die in Ops, Not Tech — Post-Launch Loop, Safety Layer & 30-Day Monitoring Plan25 min read

My academic papers, unpacked: from reward hacking to catastrophic forgetting — how the experiments were designed and what they mean for engineering.

1. 01Reward Hacking in AI Agents: Trained 60,000 Steps, the Agent Learned to Delete Tickets (6 ITSM Patterns)30 min read
2. 02Catastrophic Forgetting in LLMs: 52 Domains Fine-Tuned, the Earlier 51 Regressed — A Dual-Replay Field Report30 min read

Engineering deep dives paired with real production incidents: refund workflows, KB desync, event-loop blocking, Claude Code token attribution.

1. 01What a Real, Money-Moving L2 Refund Workflow Actually Looks Like | Workflow Deep Dive25 min read
2. 02Your KB Changed. The Search Index Didn't — Anatomy of a 9-Day Silent Desync | KB-Ops Deep Dive20 min read
3. 03Your Observability Dashboard Is Throttling the Agent It Watches — an Async Latency Postmortem20 min read
4. 04I Ran the Whole Token-Saving Playbook. The Savings Got Re-Spent.20 min read

WeChat is beta-testing an AI agent that can search, compare, order, and pay across 10M merchants. Three posts: how to get found, how to keep your customers, how to rebuild.

1. 01AI Picked the First Store and the Other Four Vanished: WeChat's New Shelf for 10 Million Merchants10 min read
2. 02When the AI Becomes the Storefront, You Decay Into a Supplier: The Relationship Moat for 10 Million Merchants10 min read
3. 03The Day WeChat's AI Ordered for Users, 10 Million Merchants' Mini-Programs Expired — A Layered Rebuild Blueprint30 min read

Curated translations of standout external writing — original source and translator always credited (not original work).

1. 01创始人手册：打造 AI 原生初创公司30分钟阅读
2. 02Fable 实战指南：找到你的未知12分钟阅读
3. 03用 Fable 5 搭一个第二大脑14分钟阅读
4. 04Loop Engineering:Karpathy 方法论,和那篇让它再快 5 倍的研究10分钟阅读
5. 05用 Fable 5 搭一个自我改进的 agent 系统:14 步,loops、动态工作流、routines15分钟阅读
6. 06四种 Loop 分别什么时候用:Claude Code 官方的 loop 入门指南8分钟阅读
7. 07反向信息悖论：在 AI 时代，先泄密的是买方9分钟阅读
8. 08用 12 步造一个会拒绝自己幻觉的研究 agent:搜索循环、来源打分、事实闸14分钟阅读
9. 09企业财务的 AI 该怎么落地:别买平台,给「中间那个人」装个后台 agent13分钟阅读
10. 10让沙子学会思考之后：Demis Hassabis 的前沿 AI 治理框架11分钟阅读
11. 11温度计与恒温器：让 eval 真正改变下一步的那一层10分钟阅读
12. 12Harness / Loop / Graph:agent 系统坏了,先分清坏在哪一层12分钟阅读
13. 13Loop Engineering 是一种模式,不是一个功能12分钟阅读
14. 14让一个 Obsidian 研究系统替你读完这周的资料15分钟阅读
15. 15记忆工程：让 agent 的记忆越用越聪明，而不是越用越臃肿13分钟阅读
16. 16Claude 5 时代的 context 工程新规则11分钟阅读