AI Agent Roles in the Disney Model
The Disney Model of creativity is over 30 years old. Its three-way division into Dreamer, Realist and Critic provides a surprisingly precise architecture for AI agent workflows: specialized roles, strict phase separation, iterative quality assurance.
Photo: Alan Fisher / New York World-Telegram & Sun, Library of Congress (Public Domain)
The Disney Model: More Than a Creativity Technique
A Disney animator is said to have once observed that there were "three different Walts": one who dreamed wildly, one who made plans, and one you didn't want in the room when you were presenting a fresh idea. Robert B. Dilts formalized this observation in the early 1990s into a process model, which he first described systematically in Tools for Dreamers (1991).
The core idea is simple: creative performance does not emerge despite, but through the separation of three thinking functions. Generating ideas and evaluating them simultaneously blocks the process. Dilts framed the three phases as guiding questions:
Dreamer
“WANT TO”
What do we want? Why? What would be possible if there were no constraints? The Dreamer thinks big, uncensored, visionary. No feasibility check, no “yes, but.”
Realist
“HOW TO”
How do we make it happen? Who does what, when, with what? The Realist does not criticize — he makes things feasible. Steps, milestones, measurable criteria.
Critic
“CHANCE TO”
What could go wrong? Who is affected? What works well today and must not be broken? The Critic improves — he does not destroy.
The decisive detail lies in the name of the third phase: Dilts did not call it “Danger to” but “Chance to.” The Critic is not the enemy of the idea — he is its chance for improvement. Every criticism is translated into a constructive question: “How can we...?”
Why the Disney Model Fits AI Agents
At first glance, applying a creativity technique from the 1990s to AI agents might look like a neat analogy. It is more than that — and for a structural reason.
The fundamental problem in complex AI workflows is well known: a single agent that simultaneously generates, plans and evaluates produces mediocrity. This applies equally to code, designs, strategy papers and campaign concepts. The LLM has no natural phase separation. It hallucinates and corrects in the same breath. It drafts brilliant concepts and retreats to safe platitudes in the very next token — because the next token is also the evaluation of the previous one.
This is exactly the problem Dilts described in human teams: mixing thinking functions blocks the process. The solution is the same in both cases — role separation through architecture, not through discipline:
- In humans: Different rooms, different times, clear phase transitions.
- In AI agents: Different system prompts, different contexts, clear handoff artifacts.
The three Dilts roles can be directly implemented as an agent architecture:
| Role | In a Workshop | As an AI Agent | Output |
|---|---|---|---|
| Dreamer | Inviting room environment, no criticism allowed | System prompt: “Think radically, no constraints.” High temperature. | 3–5 solution approaches, payoff statements, future vision |
| Realist | Practical workspace, flip charts, timelines | System prompt: “Make it feasible. Steps, resources, criteria.” | Step-by-step plan, milestones, measurable success criteria |
| Critic | Deliberately “uncomfortable” room, hard questions | System prompt: “Evaluate against criteria X, Y, Z. Be strict.” Fresh context. | Weakness list, ecology check, “How can we...?” questions |
The Critic Agent: The Heart of Quality Assurance
Of the three roles, the Critic is the most interesting for AI workflows — because it solves the problem that LLMs handle worst on their own: honestly evaluating their own work.
An LLM judging its own output is like an author reviewing their own book. The information that led to its creation still sits in context — and colors the evaluation. The solution is architectural:
Design Principles for the Critic Agent
Fresh context — no memory of the generation process
The Critic agent starts without the context of the Dreamer and Realist phases. It only knows the standardized evaluation prompt and the finished artifact. No history, no intermediate versions, no justifications. This prevents the “sunk-cost bias” that arises when an agent recapitulates its own reasoning process.
Standardized evaluation prompt — consistent across sessions
The Critic always receives the same evaluation prompt, regardless of what the Dreamer or Realist experienced in that session. This makes evaluations comparable: what receives a 75% score in session 1 has the same standard as the result in session 30.
Concrete evaluation criteria — no vague quality judgments
Not “Is this good?” but: “Evaluate against these specific criteria. Score each component. Flag suspicious areas.” The more concrete the criteria, the more useful the result. Vague prompts produce vague criticism.
Criticism as a question — Dilts' “How can we...?”
Pure negative lists are also useless in AI workflows. The Critic agent should translate every weakness into a constructive question: not “The module is poorly structured,” but “How can the module separate responsibilities more clearly?” This gives the Dreamer a concrete starting point in the next iteration.
Ecology check — what must not be broken?
Dilts' most important and most frequently forgotten point: the Critic does not only check for risks, but also identifies which positive qualities of the status quo must be preserved. In AI workflows: which existing structure, tone or consistency with the overall project must not be lost during revision?
“The Critic must ask at least one ecology question. Pure negative lists are not permitted.”
Iteration Logic: At Least Three Cycles
A single pass (Dreamer → Realist → Critic) rarely yields a solid result. The model calls for at least three complete cycles — and this is where it becomes particularly valuable for AI workflows.
The Three Passes Differ Fundamentally
Pass 1: Exploration. The Dreamer generates broadly. The Realist sorts. The Critic identifies the most obvious weaknesses and formulates “How can we...?” questions. Result: a rough artifact plus a concrete list of improvement questions.
Pass 2: Evolution. The Dreamer now works informed. It knows the Critic's questions and integrates them — without being constrained by them. It combines the best elements from pass 1 and attempts at least one new variant. The Realist sharpens the plan. The Critic evaluates with the same standardized prompt — and compares with the result from pass 1.
Pass 3: Convergence. If approaches have remained stable across passes, the process converges. If not, it iterates further. The result is a final one-pager: vision, plan, risks, ecology check and a clear go/no-go recommendation.
Practical Implementation: An Agent Tableau
The following architecture can be directly implemented as a multi-agent system — whether with LangChain, CrewAI, Anthropic's Tool Use, or a simple orchestrator script:
| Agent | Core Responsibility | Input | Output |
|---|---|---|---|
| Dreamer Agent | Open the possibility space, generate radical options | Challenge statement, context, desired benefit | 3–5 solution approaches + payoff statements + future vision |
| Realist Agent | Create implementation plan, clarify resources | Dreamer output, constraints, available resources | 1–2 prioritized concepts + step plan + success criteria |
| Critic Agent | Ensure quality, identify weaknesses, run ecology check | Realist plan (fresh context, no Dreamer knowledge) | Weaknesses + ecology check + “How can we...?” questions + go/iterate recommendation |
| Orchestrator | Manage cycles, pass artifacts, check exit criteria | Results from all three agents | Final one-pager or trigger for next pass |
Five Rules for Practice
- Strictly separate the phases. No “By the way, is this even feasible?” in the Dreamer prompt. No “And evaluate while you're at it” in the Realist prompt. Mixing is the most common mistake — in humans and AI alike.
- The Critic always starts fresh. No context from the generation phase. No “I know you worked particularly hard on section 3.” The Critic only knows the result and the evaluation criteria.
- Define concrete thresholds. Not “Is this good enough?” but measurable criteria: test coverage > 80%, no duplicates, all required components present, accessibility score met. If the threshold is not reached: back to the Dreamer — do not tinker with the existing result.
- When stuck: back to the Dreamer. If the Critic says “Iterate” three times in a row, the problem is not in the wording but in the fundamental approach. No amount of fine-tuning will help — the Dreamer needs a fresh perspective.
- Ecology questions are mandatory. The Critic must ask: what is working well right now? What will be lost through the change? In AI workflows: which consistency, style or structure must be preserved, even when content changes?
Seven Application Areas
The Disney Model with AI agents works wherever creative work needs structure. Here are seven domains — each with a concrete example of how the three roles work together.
1. Software Development
Dreamer: Architecture drafts, feature ideas, API design variants. Realist: Write code, tests, clarify dependencies. Critic: Code review for bugs, security, performance, technical debt.
2. Product Design & UX
Dreamer: Wireframes, interaction concepts, radical UI ideas without constraints. Realist: Design-system-compliant implementation, responsive variants, component library. Critic: Accessibility check, usability heuristics, brand consistency.
3. Strategy Development & Business Planning
Dreamer: Business models, market opportunities, blue-sky scenarios. Realist: Financial plan, go-to-market, resource planning, KPIs. Critic: Risk analysis, competitive counter-check, stakeholder impact.
4. Marketing Campaigns
Dreamer: Campaign ideas, target-group personas, channel mix, creative concepts. Realist: Budget, timeline, content calendar, KPIs. Critic: Brand compliance, legal review, tonal consistency across channels.
5. Content & Text Production
Dreamer: Topic ideas, perspectives, unusual angles and metaphors. Realist: Structured texts following style guide, SEO optimization, formatting. Critic: Plagiarism check, originality check, fact verification, tone.
6. Data Analysis & Research
Dreamer: Generate hypotheses, search for unexpected correlations, exploratory visualizations. Realist: Choose statistical methods, clean data, build reproducible pipelines. Critic: Methodological critique, bias check, reproducibility, significance testing.
7. Education & Curriculum Design
Dreamer: Learning objectives, innovative formats, gamification ideas, blended learning concepts. Realist: Structure curriculum, create materials, design assessments. Critic: Didactic review, learning objective achievement, accessibility, cognitive overload.
What All Application Areas Have in Common
- The Critic always starts fresh — whether evaluating code, designs or strategies. No context from the generation phase.
- A standardized Critic prompt ensures consistent evaluations across sessions. What receives a “Go” in session 1 has the same standard as session 30.
- Some artifacts need one pass, others four. The iteration logic works — even at 4+ rounds. Rounds 3 and 4 are not touch-ups, but fundamentally new approaches.
- The Dreamer gets more cautious with each iteration — if you're not careful. Solution: give the Dreamer the Critic's questions, but not the detailed scores.
- Duplications between related artifacts are the most common blind spot. The Critic must explicitly check whether variant 2 differs sufficiently from variant 1.
Conclusion: Old Method, New Application
The Disney Model is not an AI framework. It is a facilitation heuristic from the NLP tradition, inspired by observations of an animation studio. But its core principle — deliberately separating, sequencing and iterating thinking functions — solves a problem that is just as real in AI workflows as in human teams.
The three roles give an agent architecture what it otherwise lacks: structure without rigidity. The Dreamer can be bold because the Realist will sort things out afterward. The Realist can be pragmatic because the Critic will review it. And the Critic can be strict because the Dreamer will bring a fresh perspective in the next pass.
The result: not a single, perfect shot — but a process that converges through iteration. Exactly as Walt Disney (allegedly) already knew.
Sources & Further Reading
- Dilts, R.B., Epstein, T. & Dilts, R.W. (1991): Tools for Dreamers, Meta Publications — in particular Appendix H: “Well-Formedness Conditions for Evaluating New Ideas”
- Hochschule Luzern (n.d.): Disney Method — method card with zone setup and time grid
- Windauer (n.d.): Walt Disney Creativity Strategy According to Robert B. Dilts — NLP-influenced variant with meta-position and state anchors
Related Articles
AI Development 2026: From Chatbots to Physical AI
Where does AI stand in 2026? OpenAI's 5-level model, Physical AI and the EU AI Act for HR.
Read moreAI & Psychology: When AI Models Need Therapy
What happens when we confront AI models with psychological tests?
Read moreThe Evolution of AI: AGI Through Developmental Psychology
How developmental psychology models help understand the path to AGI.
Read more