Building a stable personality for an AI agent sounds abstract. But once you accept that personality is a set of priors rather than a description, an engineering question crystallises: which personality framework do you use?
Why not Big Five
Big Five is rigorous in human research, but on agents it has a fatal problem: it's descriptive, not predictive. What does a high-neuroticism, low-agreeableness person do under pressure? Big Five doesn't tell you. You get five scores, and there's no code you can write from those five numbers.
MBTI has the opposite problem — too many types, not falsifiable, types drift over a lifetime. HEXACO adds a Honesty-Humility axis to Big Five, which helps, but it's still description-shaped.
What's special about the Enneagram
The Enneagram looks like a list of nine types. It's actually a behaviour map. Each type ships with: a core motivation, a core fear, a stress direction (where it goes when overloaded), an integration direction (where it goes when supported), two flanking wings, and three instinctual subtypes (self-preservation / social / one-to-one). Twenty-seven fine-grained subtypes total, each with predictable behavioural distributions.
For an engineer, this is the only way to turn personality into code. You stop asking "is this agent's agreeableness 0.7 or 0.3" and start asking "under stress does it move to 8 or 4, when supported does it move to 7 or 1." The second question has answers. The first one doesn't.
How Ganjiang turns the Enneagram into Markdown
Inside Ganjiang's Soul Forge, the Enneagram isn't a label — it's a set of generation parameters. A 5w4 sp/sx agent's soul.md doesn't just say "I am 5w4." It automatically produces:
- agents.md prefers terse, information-dense phrasing.
- user.md emphasises respect for the user's boundaries (typical 5 behaviour).
- identity.md records that under pressure it "goes silent and retreats into research" (a 5→7 disintegration pattern).
- soul.md states its growth direction is 8 — "when supported, it shifts from research mode to action mode."
These aren't taglines. They're behavioural priors that the agent will reach for when it hits a situation no prompt covered. That's the part that decides whether "agent works under pressure" turns out to be a coin flip or a known quantity.
That's why Ganjiang chose the Enneagram — not because it's the most popular framework, but because it's the only one that gives you an executable answer to the question "what will this agent do when I'm not watching."
“Big Five gives you five scores. The Enneagram gives you twenty-seven engineerable behaviour playbooks. The latter is what agents need.”