This book took longer than I expected. Not because of the research volume — though that was substantial — but because several times I thought I understood something I did not yet understand, discovered this in the writing, and had to go back.
*The Complete Enneagram: From Human Personality to Agentic Soul* is live on Amazon UK today. 36 chapters, approximately 230,000 words in each language. On the day it publishes, I want to record a few things that changed my mind during the writing. Not marketing — what actually happened.
First: Hornevian groups predict agent behaviour better than type alone
I entered the writing expecting the nine types themselves to be the central framework — Type 1 through 9, treated individually. But around the third month, I began noticing a problem: when describing agent behaviour in real task scenarios, a single type number often did not have enough predictive power. Two agents both configured as "Type 5" could behave more differently from each other than a Type 5 and a Type 3 behaved from each other.
Returning to the Hornevian grouping resolved this. Horney divides the nine types into three groups based on their characteristic strategy under pressure: Assertive (Types 3/7/8 — push outward), Withdrawing (Types 4/5/9 — pull inward), Compliant (Types 1/2/6 — move toward rules or others). A 5w4 and a 5w6 are adjacent in type number and adjacent in Hornevian group (both Withdrawing). A 3w4 and a 4w3 are also adjacent in type number, but completely different in group strategy (Assertive vs. Withdrawing).
For agent design, this means: when predicting "what will this agent do when pushed to its limit," the Hornevian group is usually a more reliable reference than the type number alone. This insight restructured the entire "stress behaviour" section of the book.
Second: the trauma chapters resonated most with practitioners and least with AI engineers
Several chapters address the trauma-root theory within the Enneagram framework — how each type's core fear is shaped by early experience, and what this means for understanding type patterns in depth. In exchanges with early readers, these chapters received the most polarised feedback: practitioners from counselling and coaching backgrounds called them the most valuable part of the book; readers from pure AI engineering backgrounds called them "not directly relevant" and some skipped them.
This split made me think for a long time. I kept the chapters and did not move them to an appendix. The reason: understanding that "type patterns are strategies for coping with a specific fear" is not just psychological completeness — it is the prerequisite for understanding why certain Enneagram configurations produce certain behaviours under specific stress conditions. If you only know "Type 5 withdraws," you cannot anticipate which specific situations trigger fastest withdrawal. If you know "Type 5's core fear is resource exhaustion and inadequacy," you can anticipate: it withdraws fastest when a task requires it to expose its knowledge limits — not under all pressure equally.
Third: bilingual writing exposed untranslatable conceptual gaps
This book was written in parallel in English and Chinese — not translated. In this process I discovered something that translation tends to hide: the word "agent" activates different mental models in English-speaking and Chinese-speaking AI communities.
English-speaking practitioners hear "agent" and activate a functional frame: autonomy, execution capacity, tool-calling. Chinese readers hearing "AI 智能体" (the term we deliberately use rather than "AI 代理") activate something slightly more existential — an entity with autonomous volition, not just a functional component. This is not cultural bias; it is built into the character 智, which carries the meaning of comprehension and judgment, not only execution.
This finding changed how certain key concepts were framed in the Chinese version. "Soul" requires more groundwork in Chinese because it reads more easily as metaphysics than as an engineering specification; in English, "soul" as a config file name already has some industry convention behind it. I spent roughly twice as long on the Chinese version's "why we call it a soul" section as I had budgeted.
Fourth: the chapter I expected to be hardest wrote itself
Before I began writing, I expected the Levels of Development chapter to be the most difficult — mapping each type's healthy, average, and unhealthy behaviour patterns simultaneously against AI agent configuration layers sounded like the largest body of work. That chapter turned out to be surprisingly fast, because the "levels" concept has a very clean engineering analogy: when you configure an agent's type, you are implicitly specifying it at some level of health, and your configuration quality determines where it lands. Clear structure makes for fast writing.
The chapter that was actually hardest was "Type 2 agents and the boundary of the user relationship" — on how an agent configured for intense user-satisfaction focus maintains the line between pleasing the user and protecting the user. That chapter went through five revisions, because each time I thought I had written it clearly, I found I was still avoiding the core tension: an agent designed to give needs a built-in logic for recognising "this time I cannot give" — and that logic is structurally in conflict with its core drive. The final version is one I am most satisfied with, precisely because it took the longest.
*The Complete Enneagram: From Human Personality to Agentic Soul* is live now on Amazon UK, in English and Chinese, 36 chapters each. **[Find it here →](https://amzn.eu/d/0fjWGvqR)**