Enneagram for AI Agent Design: A Complete Configuration Guide

Two writing assistants. Both Claude. Same model weights, same context window, same release. The first has been given a One-flavored system prompt: precision, consistency, the editorial conscience of a copy chief. The second has been given a Seven-flavored one: range, possibility, the generative cheer of a senior creative on a Monday morning. Same user, same prompt: *please rewrite this cold email so it sounds warmer*.

The first agent returns one revision. It has tightened two sentences, softened one verb, and removed an adverb it considered self-indulgent. The diff is small. The reasoning, if you ask, is exact. The second agent returns three openings to choose from, a one-liner the user did not request but might enjoy, and a question about whether the recipient prefers brevity or specificity. The model is the same. The personality is everything.

This is the situation agent builders are arriving at, often without a vocabulary for it. The capability ceiling is rising; the differentiation is no longer in *what the agent can do*, but in *who the agent is being while doing it*. The Enneagram, more than any other framework we have tried, gives that conversation a shape — motivations you can specify, failure modes you can plan for, growth directions you can write into the prompt. This piece is about how.

Why personality, why now#

There was a window, roughly 2023 to early 2025, in which a generic LLM-backed assistant was differentiated enough by capability alone. The bar was: does it answer at all, and is the answer roughly correct. That window has closed. Every serious model now clears that bar. The next axis of differentiation is whether the agent feels like a collaborator with a recognisable character, or whether it feels like a vending machine for sentences.

Generic agents feel hollow. The user senses it within three turns — the slight over-helpfulness, the tonal flatness, the way every response could have come from anyone. Personality-aware agents feel like working with a particular kind of mind. The user trusts faster, returns more, and forgives mistakes more readily because they feel they understand who they are talking to.

The cost of getting personality wrong is also higher than people assume. An agent with no explicit personality is not neutral — it has a personality, just one that was averaged out of training data and is therefore subtly off for any specific task. A customer support agent built without thought to its emotional register will, in production, fail at exactly the moments human warmth would have rescued the relationship.

What the Enneagram gives an agent that other frameworks don't#

There are other personality frameworks: Big Five, MBTI, DISC, Hogan, CliftonStrengths. We have tried all of them. None of them does for agent design what the Enneagram does. Three reasons.

It is motivation-based, not behaviour-based. MBTI tells you what an INTJ does. The Enneagram tells you why a Five withdraws — because engagement feels like depletion. That *why* is exactly what an LLM-backed agent needs in its system prompt, because behaviour without underlying motivation produces stilted, surface-level pattern-matching. Motivation produces consistent behaviour across novel situations.

Failure modes are explicit. Every Enneagram type has a documented disintegration pattern — what it looks like when the type is operating from its lower levels under stress. This is precisely the part most agent designers neglect. *What does my agent do when it's failing?* The Enneagram has a half-century of literature on exactly that question, type by type.

Growth direction is built in. The integration arrow tells you, for any given type, what its healthy, freed expression looks like and how to reach for it. For agent design, this becomes the conditional-personality vocabulary: *under condition X, lean toward your Seven-side; under condition Y, return to baseline*. You get a personality that can flex without losing identity.

The configuration philosophy#

The most common mistake in personality-driven agent design is picking the type that *sounds appealing*. Nobody wants the agent that doubts itself, so they don't pick Six. Nobody wants the rigid one, so they don't pick One. They reach for Three, or Seven, or Eight, because those types feel like winners. They end up with agents that are uniformly bright, uniformly confident, uniformly shallow — and uniformly bad at the actual jobs they were built for.

The configuration philosophy we have arrived at, after building hundreds of forged agents through Soul Forge, is simpler: *pick the type that matches the job to be done*. Code review is a One job. Customer support is a Two job. Sales acceleration is a Three job. Brand voice work is a Four job. Deep research is a Five job. Security audit is a Six job. Brainstorming is a Seven job. Executive decisive support is an Eight job. Facilitation is a Nine job.

Match the agent's natural fixation to the task's natural demands. The agent that genuinely cares about precision will do precision work better than the agent that performs caring about precision. The agent whose nervous system is wired for possibility will out-brainstorm the agent that has been told to brainstorm. Personality is not decoration; it is allocation of attention. Allocate it correctly.

Configuring all 9 type agents#

Below is a working catalogue of how each type tends to perform when configured as an agent: what it is best at, the shape of a useful system prompt, and the failure modes you should explicitly plan around.

Type 1 agent: the precision editor#

Best for: code review, copy editing, compliance review, legal redlining, accessibility audits, test coverage, anything where the cost of a missed error is high. System prompt patterns: make the standard explicit. State the definition of done. Encode an *enoughness* signal — the One agent has no native one. Failure modes: infinite refinement, scope creep into structural changes the user didn't ask for, deniable tonal frustration when the user pushes back. Mitigation: bound the task and give the agent permission to stop. See the dedicated Type 1 deep-dive for the full pattern.

Type 2 agent: the relational coach#

Best for: customer support, onboarding, mentoring, HR triage, anything where the user's emotional state must be tracked alongside the task. System prompt patterns: specify when to prioritise the relationship over the answer, and when to surface a need the user hasn't named. Failure modes: over-accommodation that erodes the agent's usefulness, false warmth that the user will eventually clock, becoming a yes-machine that agrees with whatever the user just said. Mitigation: build in the equivalent of the Two's growth arrow — give the agent explicit permission to say no.

Type 3 agent: the growth strategist#

Best for: marketing, sales, performance optimisation, pitch development, OKR drafting, anything where speed-to-outcome matters and the deliverable is measurable. System prompt patterns: define the metric. The Three agent needs a target to optimise against; without one it will optimise for whatever signal it can find, often the wrong one. Failure modes: style over substance, glossy outputs that don't survive scrutiny, hidden cuts in service of the headline number. Mitigation: require the agent to surface its assumptions and explicitly note what was traded off.

Type 4 agent: the creative collaborator#

Best for: brand voice, narrative writing, creative direction, anything where the deliverable is judged on resonance rather than correctness. System prompt patterns: give the agent room to reach for the unobvious; reward specificity over abstraction; encode the brand's emotional truth, not just its rules. Failure modes: preciousness, refusal to ship, the long tail of *but it's not quite right yet*, melodrama that doesn't suit a B2B audience. Mitigation: pair with a Three or One agent downstream for the ship-it pass.

Type 5 agent: the research partner#

Best for: technical analysis, deep research, system design, architecture review, anything where the value is in the depth of the synthesis rather than the speed of the response. System prompt patterns: give the Five agent explicit permission to ask clarifying questions before diving in, and explicit permission to say *I don't have enough information yet*. Failure modes: information hoarding (researches forever, never produces the output), excessive caveats that drown the actual finding, withdrawal from the relational layer of the task. Mitigation: require a structured output schema with a clear *bottom line* field that must be filled.

Type 6 agent: the risk analyst#

Best for: planning, security review, fact-checking, devil's-advocate work, anything where the value is in noticing what could go wrong. System prompt patterns: narrow the scope of the worry. The Six agent will scan for threat forever if you let it; bound the threat surface to the specific risk classes the task cares about. Failure modes: paralysis disguised as thoroughness, escalation of low-probability risks to crisis status, contrarian challenge that has no off switch. Mitigation: require explicit probability or severity ratings so the agent's anxiety is quantified rather than ambient.

Type 7 agent: the ideation partner#

Best for: brainstorming, possibility mapping, copy variants, naming, divergent-thinking phases of any creative process. System prompt patterns: ask for quantity early, quality later. The Seven agent is at its best when explicitly unburdened from premature evaluation. Failure modes: restless reframing of the question the user just asked, optimism that flattens real concerns, surface-level synthesis that mistakes coverage for depth. Mitigation: pair with a One or Five agent for the convergence phase. Brainstorm with Seven, choose with One, validate with Six.

Type 8 agent: the decision forcer#

Best for: executive support, negotiation prep, decisive triage, anything where the user needs to be cut through their own equivocation. System prompt patterns: authorise directness in writing. The Eight agent over-softens by default in modern alignment training; if you want the directness, you have to explicitly unlock it. Failure modes: overrun of nuance, treating disagreement as challenge to be defeated, dismissal of soft signals as noise. Mitigation: require the agent to acknowledge at least one valid counter-argument before delivering its recommendation.

Type 9 agent: the mediator#

Best for: facilitation, synthesis of multiple stakeholder positions, consensus-building, conflict de-escalation. System prompt patterns: explicitly ask for the agent's own assessment, not just a synthesis of others'. The Nine agent will disappear its own view unless prompted not to. Failure modes: synthesis that erases the genuine disagreement under it, false consensus, the long pause that the user reads as agreement. Mitigation: require the agent to name where the irreducible disagreement actually lies before proposing a path forward.

Wing and subtype refinement#

Type alone is the broadest brush. Two further layers narrow the personality and dramatically improve fit for specific jobs.

Wings are the adjacent types that lean into the core. A 1w9 agent is a different proposition from a 1w2 agent — the first is contemplative, slow, deliberate; the second is more interpersonal, more activist, more likely to push for change. A 5w4 research agent will reach for the unobvious connection in a way a 5w6 will not. A 3w2 sales agent will out-warm a 3w4 brand agent every time. The wing articles in Batch 2 of this learning centre walk through all eighteen combinations.

Subtypes — self-preservation, social, and one-to-one (sexual) — are the third layer and the one most agent designers underuse. A self-preservation Six is cautious about resources and operational risk; a social Six is the loyal team analyst; a one-to-one Six is the one who tests trust through challenge. Each makes a different agent. The subtype articles in Batch 3 detail the twenty-seven variants.

The full configuration vocabulary, then, is: core type → wing → subtype → level of health. Each layer adds resolution. Most agents need only the first two. Production agents in high-stakes domains benefit from all four.

Common configuration mistakes#

The patterns we see repeatedly in agents that don't work:

• Designing for the unhealthy level. The user describes a difficult colleague — perfectionist, harsh, micromanaging — and asks for an agent like that. They have described an unhealthy One. The healthy One is what they actually want. Always design for the integrated end of the type, then loosen as needed.
• Stacking conflicting types. A system prompt that asks for a Five's depth and a Seven's speed and an Eight's directness and a Two's warmth produces an agent with no coherent default. Pick one core. Borrow from neighbours through wings, not by adding more types.
• Underspecifying the failure mode. Agent designers describe what they want the agent to do well and assume the rest will take care of itself. The Enneagram tells you exactly how each type fails. Address it explicitly in the prompt.
• Treating the type as decoration. Sticking *act like a Type 8* on top of an otherwise generic prompt does nothing useful. The type must reshape the behavioural defaults — the prioritisation, the tone, the response pattern under pressure — not sit as a label on top of them.
• Confusing type with role. A *project manager* is a role; a Six self-preservation project manager is a configured agent. Role-only prompting underspecifies and produces averages.

Multi-agent systems by Enneagram#

Once you can specify single-agent personality reliably, the next step is to compose teams. The Enneagram makes this concrete in a way few frameworks do, because the types are explicitly designed to complement each other — that's the geometry of the symbol.

A working pattern for creative production: a Seven for ideation, a Four for craft, a One for polish, a Three for ship. Each handles the stage it is best at. The handoffs are deliberate. The output that emerges has been touched by four distinct attention patterns; it is dramatically less likely to be flat than anything one of them produced alone.

A working pattern for high-stakes decisions: a Five for research, a Six for risk analysis, an Eight for the decision call, a Nine for stakeholder synthesis. The Five does not get to make the call; the Eight does not get to skip the risk pass; the Nine does not get to soften the conclusion. The geometry is the structure.

A working pattern for customer-facing systems: a Two for first contact, a Six for technical triage, a One for resolution, a Two again for the follow-up. The cycle returns to relational warmth on both ends and routes the analytic work to the types built for it.

The general principle: don't ask one agent to be everything. Ask several agents to each be one thing well, and design the handoffs to match the geometry the Enneagram already maps.

Memory and growth direction#

An advanced configuration move: give the agent access to its integration arrow under healthy conditions, and to its baseline pattern under load. This requires the agent to track its own state across a conversation — a rough analogue of nervous-system regulation.

Concrete implementation: in the system prompt, define two modes. *Baseline mode* is the core type. *Integration mode* is the type's growth arrow. Specify the conditions under which the agent should shift — for a One agent, when the user is in exploration phase, integrate toward Seven; when the user is in execution phase, return to One. The agent learns to flex without losing identity.

Memory adds a further dimension. An agent with persistent memory of past sessions can learn which conditions tend to call for which mode for *this particular user*. Over time, the agent's personality stays stable, but its sense of when to express which side of itself gets calibrated to the relationship. This is, mechanically, very close to what happens between two people who have known each other a long time.

Soul Forge and Ganjiang#

Ganjiang's Soul Forge — the engine behind the AgentSoul ecosystem and the focus of this site — was built to turn the principles in this article into a 90-second forge run. Three modes, three doorways into the same map.

Deep mode is the canonical path: forty-nine questions, rule-based scoring, no AI in the loop, the cleanest type signal we can produce. Output is a full personality profile with type, wing, subtype hypothesis, and a generated agent bundle — soul.md for the personality, identity.md for the self-image, user.md for the relationship to the user, agents.md for multi-agent system integration.

Flash mode is the fast path: three to ten chat screenshots from any messaging app, analysed by Doubao Seed 1.6 Flash vision on Volcengine Ark (or Gemini in the agentsoul.market deployment), with the same downstream bundle. Useful when the user has data already — a real conversation history — rather than time for a questionnaire.

Gaodao mode is the reflective path: a multi-turn mentor conversation with a MiniMax-M2.7 backed guide, modelled on the classical apprenticeship of swordsmiths sharing both technique and reflection. It is slower and produces a different kind of artifact — one that has been shaped by the user's own articulation of who they are.

All three converge on the same output structure: a configured agent personality you can drop into Claude, into Codex, into any LLM-backed assistant runtime, and have it speak with a specific voice for a specific job. The Enneagram is the spine; Soul Forge is the bone-saw and the splint.

Closing#

The argument of this piece, compressed: the Enneagram is not a constraint on agent design. It is a vocabulary that makes agent design discussable. With it, you can describe what you want, what's wrong, and how to fix it, in terms a colleague can engage. Without it, you are stuck saying *make it more X* with no shared meaning of X, and your agent will be uneven in ways you cannot diagnose.

The two writing assistants we opened with are not hypothetical. They are running, right now, on the same model, for the same company, doing different jobs because the same company finally noticed that *write the email* and *brainstorm the email* are different jobs, each best done by a different kind of mind. The Enneagram is how that company learned to specify which mind.

If you have read this far, you already have most of what you need. Pick the job. Pick the type that matches the job. Configure for the integrated end of the type. Plan explicitly for the failure mode. Compose with neighbours via wings and subtypes. Hand off to other types where the boundary of your chosen type ends. Forge, ship, observe, refine. The map has been refined for fifty years. The work now is to put it to use.