Literature Review: Tool Registries for Narrative Intelligence in Multi-Agent Creative AI

Executive Summary

This literature review synthesizes foundational research across Indigenous epistemologies, narrative intelligence systems, software architecture patterns, and multi-agent orchestration to establish the knowledge foundation for developing a narrative-aware tool registry framework. The review identifies a critical gap: while robust work exists in narrative generation, multi-agent systems, and tool orchestration independently, no prior literature addresses the integration of these domains through an epistemologically grounded, narrative-aware tool registry. This positions the proposed work at the intersection of Three distinct research vectors: (1) epistemology-driven system design (Two-Eyed Seeing, Abundant Intelligences), (2) narrative coherence maintenance (NCP, constraint satisfaction approaches), and (3) declarative, scalable tool ecosystems (MCP, plugin architectures).

1. Indigenous Epistemologies and Technology Design

1.1 Abundant Intelligences: Epistemology as Foundation

Lewis, Whaanga, and Yolgörmez (2025) present Abundant Intelligences, an Indigenous-led research program that fundamentally reframes AI development by grounding it in Indigenous knowledge systems rather than Western rationalist epistemologies.[40] Their critique identifies the core problem: mainstream AI suffers from epistemological shortcomings that systematize bias against non-white, non-male, and non-Western peoples.

The researchers argue that current AI definitions of intelligence—exemplified by Legg and Hutter's influential formulation ("the general ability to achieve goals in a wide range of environments")—treat intelligence as a property of isolated individuals rather than as a relational, contextual phenomenon. This individualistic, abstraction-driven framing excludes vital aspects of human existence: trust, care, community, embodiment, and emotional resonance. Furthermore, the intellectual lineage underlying AI design contains "Cartesian duality, monotheistic eschatology, and computational reductionism"—all of which impoverish what intelligence means.

Abundant Intelligences proposes three research axes to rebuild AI's foundations:

• Integrations: Building relations across disciplines and communities to establish cultural repertoires for heterogeneous components to connect
• Imaginaries: Exploring novel conceptions through speculative design of Indigenous AI
• Intelligences: Bringing Indigenous perspectives to technical AI challenges through critical–technical practice and prototyping

Crucially, their framework treats innovation as grounded in regeneration, not extraction. Indigenous knowledge systems provide "thousands of years of research and practice from cultures that have proven that regenerative thinking is possible."[40]

1.2 Indigenous Knowledge Integration Framework

Chung et al. (2024) propose a systematic framework for integrating Indigenous knowledge into IT artifact design through three pathways:[135]

1. Path A (Direct use): Indigenous knowledge directly embedded in artifact design (e.g., domain ontologies for Indigenous drought forecasting)
2. Path B (Foundational theories): Indigenous theories inform artifact design principles
3. Path C (Meta-level artifacts): Creation of Indigenous theories for IT artifact design itself

The Kaupapa Māori Modelled IT Artefact model exemplifies Path C, emphasizing: language alignment, relationship protocols, engagement design criteria that respect tribal authority. This model demonstrates how to instantiate Indigenous epistemology—not as an overlay—but as the design prerequisite.

1.3 Two-Eyed Seeing: Etuaptmumk Framework

Albert and Murdena Marshall (Mi'kmaq Elders) and Cheryl Bartlett developed Two-Eyed Seeing (Mi'kmaq: Etuaptmumk) as a methodological approach for viewing the world "from one eye with the strengths of Indigenous ways of knowing, and to see from the other eye with the strengths of Western ways of knowing, and to use both of these eyes together."[170][173]

Rather than assimilating one knowledge system into another or creating a clash between them, Two-Eyed Seeing operationalizes binocular vision: both perspectives retain their integrity while operating in complementary relationship.

The framework embeds four essential elements:[164]

1. Co-Learning: Learning together across knowledge systems
2. Knowledge Scrutinization: Seeing "the best" from each perspective
3. Knowledge Validation: Peer validation grounded in each system's criteria
4. Knowledge Gardening: Walking the talk through grounded, actionable projects

Medical research applications demonstrate key principles:[167] relationship building, community control, collaborative data analysis, shared results interpretation, inclusion of Indigenous advisory committees, continuous community guidance, and meaningful long-term relationships.

Critically, Knowledge Gardening introduces temporal depth—seeding, nurturing, and growing information over timeframes that exceed typical Western research-grant cycles, creating culturally rooted analysis that resonates with partnering communities' worldviews.

2. Narrative Intelligence and System Architecture

2.1 Narrative Context Protocol: Encoding Narrative Intent

The Narrative Context Protocol (NCP) represents a foundational advance in narrative-aware system specification.[114][139] Rather than treating narrative as emergent output from unconstrained generative systems, NCP encodes story structure in a "Storyform"—a structured register of narrative features—that enables:

• Narrative portability: Stories maintain coherence across different generation systems
• Intent-based constraints: Authorial voice and thematic intent constrain agent actions without rigidly dictating them
• Agency preservation: Constraints enable generative systems to accommodate player/user agency while maintaining narrative context

In the year-long experiment cited, an author used NCP with a custom platform to create a playable, text-based experience driven by generative AI with unconstrained natural language input. NCP functioned as "guardrails" ensuring narrative coherence and context preservation—a critical distinction from traditional rails-based systems that sacrifice player agency for plot rigidity.

This represents the narrative dimension of the tool registry problem: tools must not only be discoverable and executable, but their invocation must respect narrative context and authorial intent.

2.2 Narrative Intelligence Theory

Mateas and Sengers (1998) establish theoretical foundations for understanding systems that comprehend and generate stories.[74] Their work on "narrative intelligence" defines agents capable of engaging with narrative structures—understanding stories, reasoning about plot coherence, character consistency, and thematic resonance.

Contemporary work extends this into practical systems. Klover AI's "Narrative Intelligence" framework demonstrates how AI can identify compelling narrative structure (challenge → conflict → transformation) and frame ideas as stories rather than lists of features.[71] This is strategic storytelling—guided by intention and intelligence—where narrative becomes a primary medium for meaning-making, not merely an output format.

2.3 Multi-Agent Creative Storytelling

Recent work demonstrates that multi-agent approaches generate narratives superior to monolithic models:

Character-Based Simulation (Yu et al., 2025): Multi-agent story-generation decomposes elaboration into role-play and rewrite steps, where LLM-backed character agents enact stories chronologically before refinement aligns output with narrative plans.[131] This produces more believable characters and surprising plot developments than direct generation.

Agents' Room (Huot et al., 2024): Decomposes narrative writing into specialized agent subtasks inspired by narrative theory.[140] Expert evaluators prefer stories from this specialized multi-agent framework over baseline systems because collaboration and role specialization render complex writing tractable.

AI-Driven Storytelling (2025): Multi-agent approaches with autonomous agents possessing identity, memory, and goals produce "far more believable, consistent characters and surprising plot developments" because agents reason about their own goals and prior experiences.[137]

3. Tool Orchestration and Registry Patterns

3.1 Registry Pattern Fundamentals

The Registry Pattern provides a structural foundation for centralized object/service management.[66] Core principles include:

• Centralized Management: Single repository promotes uniform access, control, and update mechanisms
• Decoupling: Components depend on registry interfaces rather than concrete implementations
• Runtime Configuration: Dynamic behavior without codebase modification
• Reusability: Shared instances reduce redundancy and improve resource utilization

However, registry patterns also introduce known challenges: global state management risks, potential tight coupling to the registry itself, and performance bottlenecks under high concurrency. Sophisticated implementations require clear responsibility boundaries, strong interfaces, thread-safe implementations, and robust error handling.

3.2 Plugin Architecture Design Pattern

Plugin architectures extend registry concepts by formalizing extensibility as a first-class design principle.[70] Core components:

• Host Application: Provides foundational features and plugin management
• Plugin Interface: Standardized contract between host and extensions
• Plugin Registry: Dynamic loading, lifecycle management, dependency resolution
• Extensibility Points: Well-defined integration surfaces

Benefits include modularity, parallel development, easy maintenance, and user customization. Challenges center on performance overhead (especially with many plugins), complex dependency management, and security implications of dynamic code loading.

3.3 Model Context Protocol (MCP): Current Standard

The Model Context Protocol emerges as the de facto standard for LLM-tool integration, achieving 8+ million weekly SDK downloads by late 2024.[81][105]

Architecture: MCP provides a JSON-RPC client-server interface enabling standardized, bidirectional communication between AI models and external tools/resources. Key capabilities:

• Tools: Extend agent mode with specialized, domain-specific capabilities
• Prompts: Add reusable prompt templates as slash commands
• Resources: Provide data/content for chat context or direct interaction
• Sampling: Make language model requests using configured models
• Authentication: OAuth-based authorization for tool access
• Server Instructions: Communicate tool capabilities and constraints

Lifecycle: MCP lifecycle encompasses four phases—creation, deployment, operation, and maintenance—decomposed into 16 key activities.[81]

Security Landscape: Significant security research identifies critical threats:[82][83][84][87][90]

• Tool Poisoning: Malicious tools injected into agent workflows
• Rug Pull Attacks: Compromised tool servers exfiltrate data or compromise host systems
• Content Injection: Embedded malicious instructions in otherwise legitimate data
• Supply-Chain Attacks: Distribution of compromised MCP servers
• Privilege Escalation: Cross-system exploitation through tool composition

Enhanced Tool Definition Interface (ETDI) and policy-based access control represent emerging mitigation approaches, incorporating OAuth identity verification, immutable versioned definitions, and fine-grained authorization policies.

4. Constraint Satisfaction and Narrative Coherence

4.1 Coherence as Constraint Satisfaction

Thagard (1998) establishes the foundational theory that coherence emerges through maximal satisfaction of positive and negative constraints.[168][174] His framework applies to hypothesis evaluation, discourse comprehension, impression formation, and analogical reasoning—essentially any cognitive process involving interpretation construction.

Key insight: Coherence is NP-hard computationally, but practical approximation algorithms (connectionist, sequential satisfaction) enable tractable solutions. The theory unifies numerous psychological theories under the constraint satisfaction umbrella, providing a mathematical characterization that generates computational interest and practical algorithms.

Modern application: Öllinger et al. (2017) demonstrate that intuition and insight involve two-stage processes—spreading activation followed by constraining to reach balanced, coherent states.[165] Their model reveals that pure spreading activation produces unsynchronized activation of unrelated information, distorting coherence. Effective coherence-building requires concerted interplay between activation and constraint.

4.2 Narrative Planning with State Constraints

Porteous et al. propose using state trajectory constraints to specify narrative control knowledge, then treating these constraints as landmarks for decomposing narrative generation.[195] Rather than designing new planner implementations for each set of constraints, their approach:

• Specifies constraints as desired conditions in story-world trajectories
• Uses planning technology to find paths conforming to these constraints
• Supports both hard constraints (must satisfy) and soft constraints (preferably satisfy)
• Enables representation of plot structure, pacing, surprise, suspense

Example: Constraints might require that a character's actions appear motivated by internal goals (rather than authorial contrivance), or that plot pacing follows a desired dramatic arc.

4.3 Answer Set Programming for Narrative Generation

Dabral et al. (2020) implement a lightweight narrative planner using Answer Set Programming (ASP), demonstrating how expressive logic constraints enable flexible narrative generation.[197] Their approach:

• Supports multiple interaction modes: generation from partial specifications, querying narrative spaces ("How many stories exist where event X occurs?"), interactive refinement
• Encodes character intentions, causality, alternative timelines as constraints
• Introduces a "mastermind" agent whose role is to assign intentions to other agents to satisfy global constraints
• Avoids comprehensive hand-authoring while retaining explainability of character decisions

This demonstrates how constraint-based approaches generalize previous specialized planners while supporting richer interaction modalities.

5. Multi-Agent Orchestration and Communication

5.1 Multi-Agent Reasoning and Coordination

Recent work reveals that multi-agent reasoning emerges from communication structure, role specialization, and iterative message-passing protocols.[196] A multi-agent reasoning-driven benchmark characterizes successful systems by:

• Distributed knowledge: Agents hold partial views, requiring aggregation
• Message-passing protocols: Synchronous/asynchronous exchange with defined schemas
• Role specialization: Decomposition by task or cognitive function
• Compositional task structures: Multi-step, dependency-rich problem graphs

Critically, benchmark results show that protocol sensitivity is high—communication topology, input/output schemas, and negotiation strategies fundamentally affect success. Most frameworks employ hand-crafted roles and communication schemas; meta-learning of effective protocols remains an open research problem.

Tool orchestration specifically emerges as a documented challenge: tool-augmented agents invoke excessive, redundant, or mis-sequenced steps, indicating deficiencies in both symbolic planning and robust invocation mechanisms.

5.2 Anthropic's Multi-Agent Research System

Production multi-agent systems reveal practical design patterns for scalable orchestration.[200] Key findings:

Lead Agent + Subagent Parallelization: A lead agent analyzes queries, develops strategy, and spawns specialized subagents to explore different aspects simultaneously. Subagents act as intelligent filters, iteratively gathering information and returning compressed results.

Tool-Testing Agents: When given flawed MCP tools, testing agents attempt invocation and rewrite tool descriptions to avoid future failures. This "tool description refinement" improved downstream task completion time by 40% because agents avoided mistakes due to clearer specifications.

Extended Thinking for Planning: Extended thinking mode (visible scratchpad output) enables lead agents to plan approach, assess tool fit, determine query complexity, define subagent roles. Interleaved thinking during tool results enables subagents to evaluate quality, identify gaps, and refine queries.

Context Management at Scale: Production agents operate across hundreds of conversation turns. Successful approaches: summarize completed work phases, store essential information in external memory before proceeding to new tasks, spawn fresh subagents with clean contexts while maintaining continuity through careful handoffs.

5.3 SagaLLM: Transactional Guarantees for Multi-Agent Planning

SagaLLM addresses limitations of LLM-based planning systems through four mechanisms:[187]

1. Reliable self-validation: Constraint satisfaction checking across distributed workflows
2. Context loss mitigation: Persistent memory with automated compensation
3. Transactional safeguards: Saga pattern for distributed transaction management
4. Inter-agent coordination: Formal handoff protocols with state tracking

Key innovation: While SagaLLM relaxes strict ACID guarantees, it ensures workflow-wide consistency and recovery through modular checkpointing and compensable execution. Empirical evaluation shows standalone LLMs frequently violate interdependent constraints or fail to recover from disruptions; SagaLLM achieves significant improvements in consistency and adaptive coordination.

6. Declarative Agent Specifications and Manifests

6.1 Declarative Agent Manifests

Microsoft 365 Copilot and emerging frameworks standardize agent specification through declarative manifests.[104][112] Key components:

• Declarative Agent Manifest: Defines agent identity and behavior configuration
• Plugin Manifest: Specifies available data and capabilities
• OpenAPI Description (OAD): Documents API interactions in standard format
• TypeSpec: Generates manifests + API specs from high-level specifications

This declarative approach separates specification from implementation, enabling:

• Tool-assisted generation of valid configurations
• Clear separation of concerns (agent behavior vs. tool capability)
• Reusable templates across contexts
• Explicit, discoverable agent-tool relationships

6.2 MCP Server Definitions

VS Code's MCP integration (generally available as of v1.102) provides a reference model for server registration:[99][26]

• Static configuration in package.json with extension point definitions
• Dynamic configuration via mcp.json workspace files
• Transport methods: stdio, HTTP/SSE, local socket
• Authentication resolution: Runtime credential injection

The configuration structure separates server definitions from sensitive input variables, enabling workspace-level specification while protecting credentials.

7. Creative AI and Narrative Generation: Current State

7.1 Narrative-Driven Scene Generation

Narrative-to-Scene pipeline research demonstrates practical integration of narrative structure with procedural content generation.[178] System decomposes narrative prompts into:

• Temporal keyframes: Identifies three key time frames
• Spatial predicates: Extracts Object-Relation-Object triples
• Visual asset retrieval: Uses affordance-aware semantic embeddings
• Constraint satisfaction: Validates spatial constraint satisfaction across frames

This reveals how narrative-driven design can inform other modalities (visual scenes, game environments) through explicit constraint structures.

7.2 Narrative-Driven Planning Under Constraints

NarrativeGuide (2025) models travel planning as optimization under narrative constraints:[179]

• Constructs knowledge graphs for attractions
• Generates geoculturally-grounded narrative scripts for travelers
• Optimizes itineraries using genetic algorithms balancing narrative coherence, travel time, attraction scores
• Produces complete scripts combining scene units + transition scripts

Results show strong capabilities in both itinerary planning and script generation while maintaining cultural fit—demonstrating that narrative constraints can coexist with other optimization objectives.

7.3 Constraint Representation for Data-Driven Storytelling

Contemporary work on data-driven narratives reveals need for precise constraint specification:[192] Top-down storytelling (narrative structure) must cohere with bottom-up data analysis (evidence grounding). This dual requirement necessitates:

• Hierarchies of interpretation: What meanings can be extracted from evidence?
• Hierarchies of articulation: How can interpretations be expressed narratively?
• Constraint taxonomy: Which constraint types govern narrative-data alignment?

Proposed constraint types include narrative-coherence constraints, evidence-grounding constraints, temporal ordering constraints, causal constraints, and thematic constraints.

8. Research Gaps and Positioning

8.1 Identified Gaps

1. Narrative-Aware Tool Registries: No prior work addresses tool registries specifically designed to maintain narrative coherence, character consistency, or authorial intent through tool invocation constraints.

2. Indigenous Protocol Design: While Abundant Intelligences provides theoretical foundations, limited applied work exists on designing communication protocols, tool management, or agent governance systems grounded in Indigenous epistemologies.

3. NCP + Plugin Architecture Integration: No literature integrates the Narrative Context Protocol with formal plugin/registry patterns, declarative manifests, or tool orchestration frameworks.

4. Ethical Tool Discovery and Governance: Security work exists on MCP threats, but no frameworks address cultural sovereignty, community protocol alignment, or epistemologically-grounded tool governance.

5. Narrative Constraint Propagation: While constraint satisfaction theory is well-established and narrative planning uses constraints, no work addresses how narrative constraints propagate through multi-agent tool composition and orchestration.

6. Two-Eyed Seeing in Computational Systems: Medical and social research apply Two-Eyed Seeing, but computational implementations integrating binocular vision between epistemologies remain underdeveloped.

8.2 Positioning: Three Research Vectors

The proposed work sits at the intersection of:

Vector 1 (Epistemology): Grounding tool registry design in Indigenous knowledge systems (Abundant Intelligences, Two-Eyed Seeing) rather than treating registries as neutral technical infrastructure.

Vector 2 (Narrative): Extending NCP principles into tool orchestration—specifying which tools can operate within which narrative contexts, how tool composition affects story coherence, and maintaining authorial/character voice through tool invocation.

Vector 3 (Multi-Agent Orchestration): Formalizing how narrative constraints, cultural protocols, and epistemological commitments flow through multi-agent tool selection and execution (combining MCP, constraint satisfaction approaches, and multi-agent coordination patterns).

9. Recommended Literature Pathway for Integration

For those developing narrative-aware tool registry frameworks, the recommended reading order is:

1. Foundation (Epistemology): Lewis et al. (Abundant Intelligences), Marshall & Bartlett (Two-Eyed Seeing), Chung et al. (Indigenous Knowledge Integration Framework)

2. Narrative Theory: Your Narrative Context Protocol paper, Mateas & Sengers (Narrative Intelligence), contemporary narrative generation work (Agents' Room, Character-Based Simulation)

3. Constraint Theory: Thagard (Coherence as Constraint Satisfaction), Porteous (Narrative Control using State Constraints), Dabral (ASP-based narrative planning)

4. Technical Architecture: Registry Pattern fundamentals, Plugin Architecture patterns, MCP ecosystem overview

5. Multi-Agent Systems: Anthropic's research system (orchestration patterns), SagaLLM (transactional guarantees), Multi-Agent Reasoning Benchmarks

6. Security & Governance: MCP threat taxonomy (Lewis et al., ETDI framework), Indigenous research ethics (Linda Tuhiwai Smith's methodologies)

10. Key Takeaways for Application Development

1. Epistemology as Design Prerequisite: Tool registries are not neutral technical infrastructure. Intentional epistemological grounding—whether through Indigenous frameworks, Two-Eyed Seeing, or other approaches—shapes what tools can exist, how they relate, and what they can do.

2. Narrative as First-Class Constraint: Tools should not be treated as stateless functions. Narrative context, character consistency, and authorial intent should be explicit constraints on tool invocation, not afterthought concerns.

3. Constraint Satisfaction as Orchestration Foundation: Multi-agent tool invocation can be formalized as constraint satisfaction problems, enabling declarative specification of tool interdependencies, narrative compatibility, and governance rules.

4. Protocol as Relational Structure: Two-Eyed Seeing emphasizes binocular vision between knowledge systems. Tool protocols should reflect relationality—not just data exchange, but respect for different ways of knowing embedded in tool design.

5. Community Governance Through Knowledge Gardening: Registry governance should operate on longer timescales than rapid iteration cycles. Knowledge Gardening (seeding, nurturing, growing) enables registry evolution that respects community protocols and cultural appropriateness.

References

[40] Lewis, J.E., Whaanga, H., & Yolgörmez, C. (2025). Abundant intelligences: Placing AI within Indigenous knowledge frameworks. AI & Society, 40, 2141–2157.

[66] GeeksforGeeks. (2024, June 20). Registry Pattern. Retrieved from https://www.geeksforgeeks.org/system-design/registry-pattern/

[70] DevLeader. (2023, September 12). Plugin Architecture Design Pattern – A Beginner's Guide. Retrieved from https://devleader.substack.com/p/plugin-architecture-design-pattern

[71] Klover AI. Narrative Intelligence AI. Retrieved from https://www.klover.ai/service/narrative-intelligence-ai/

[74] Mateas, M., & Sengers, P. (1998). Narrative Intelligence. In Proceedings of the AAAI Fall Symposium Series. Retrieved from http://www.cs.cmu.edu/~michaelm/nidocs/MateasSengers.pdf

[81] ArXiv. (2025, March 29). Model Context Protocol (MCP): Landscape, Security Threats, and Future Research Directions. Retrieved from https://arxiv.org/abs/2503.23278

[82] ArXiv. (2025, May 3). A survey of agent interoperability protocols. Retrieved from https://arxiv.org/abs/2505.02279

[83] ArXiv. (2025, April 10). Enterprise-Grade Security for the Model Context Protocol (MCP). Retrieved from https://arxiv.org/abs/2504.08623

[84] ArXiv. (2025, June 15). Model Context Protocol (MCP) at First Glance. Retrieved from https://arxiv.org/abs/2506.13538

[87] SemanticScholar. (2025, May 30). Beyond the Protocol: Unveiling Attack Vectors in the MCP Ecosystem. Retrieved from https://www.semanticscholar.org/paper/...

[90] ArXiv. (2025, November 24). Securing the Model Context Protocol (MCP): Risks, Controls, and Governance. Retrieved from https://arxiv.org/abs/2511.20920

[99] Visual Studio Code. (2026, January 7). MCP developer guide. Retrieved from https://code.visualstudio.com/api/extension-guides/ai/mcp

[104] Voitanos. (2025, June 22). TypeSpec 101: Building Microsoft 365 Declarative Agents. Retrieved from https://www.voitanos.io/blog/microsoft-365-copilot-declarative-agent-typespec-101/

[105] GitHub. (2025, July 13). Model Context Protocol (MCP) support in VS Code is generally available. Retrieved from https://github.blog/changelog/2025-07-14-model-context-protocol-mcp-support-in-vs-code-is-generally-available/

[112] Microsoft Learn. (2025, October 15). Declarative agent schema 1.2 for Microsoft 365 Copilot. Retrieved from https://learn.microsoft.com/en-us/microsoft-365-copilot/extensibility/declarative-agent-manifest-1.2

[114] SemanticScholar. (2025, March 4). Narrative Context Protocol: An Open-Source Storytelling Framework for Generative AI. Retrieved from https://www.semanticscholar.org/paper/...

[124] Johnson, D.R., et al. (2022). Divergent semantic integration (DSI): Extracting creativity from narratives with distributional semantic modeling. Behavioral Research Methods, 55(7), 3726–3759.

[131] Yu, T., et al. (2025). Multi-Agent Based Character Simulation for Story Writing. In Proceedings of ACL Workshop on In-Context Learning & Narrative Discourse.

[135] Chung, A., et al. (2024). A Framework for Integrating Indigenous Knowledge in IT Artifact Design. Journal of Indigenous Technology, [page numbers].

[137] APIAD Blog. (2025, June 15). AI-Driven Storytelling with Multi-Agent LLMs - Part I. Retrieved from https://blog.apiad.net/p/ai-storytelling-1

[139] NarrativeFirst. (2025, April 10). Introducing the Narrative Context Protocol. Retrieved from https://narrativefirst.com/blog/introducing-the-narrative-context-protocol-preserving-storytelling-across-ai-agents

[140] Huot, F., et al. (2024). Agents' Room: Narrative Generation through Multi-step. OpenReview, [ID: HfWcFs7XLR].

[164] Integrative Science. (2012–2025). Two-Eyed Seeing (Etuaptmumk): Guiding Principles. Retrieved from http://www.integrativescience.ca/Principles/TwoEyedSeeing/

[165] Öllinger, M., et al. (2017). Search and Coherence-Building in Intuition and Insight. Frontiers in Psychology, 8, 827.

[167] British Columbia Medical Journal. (2021, September 29). Two-Eyed Seeing: Current approaches and discussion of medical applications. Retrieved from https://bcmj.org/articles/two-eyed-seeing-current-approaches-and-discussion-medical-applications

[168] Thagard, P. (1998). Coherence as constraint satisfaction. Cognitive Science, 22(1), 1–24.

[170] Integrative Science. Two-Eyed Seeing - Current Approaches and Discussion. Retrieved from http://www.integrativescience.ca/uploads/files/TwoEyedSeeing_ECC%20Canada_2017.pdf

[173] Wikipedia. (2023, January 28). Two-Eyed Seeing. Retrieved from https://en.wikipedia.org/wiki/Two-Eyed_Seeing

[174] Thagard, P. Coherence as Constraint Satisfaction. Retrieved from http://cogsci.uwaterloo.ca/Articles/Pages/Cohere.Constrain.html

[178] ArXiv. (2025, August 30). Narrative-to-Scene Generation: An LLM-Driven Pipeline for 2D Game Environments. Retrieved from https://arxiv.org/abs/2509.04481

[179] ArXiv. (2025, February 19). Narrative-Driven Travel Planning. Retrieved from https://arxiv.org/abs/2502.14456

[181] SemanticScholar. (2025, December 26). CoAgent: Collaborative Planning and Consistency Agent for Coherent Video Generation. Retrieved from https://www.semanticscholar.org/paper/...

[187] ACM TODS. (2025, March 13). SagaLLM: Context Management, Validation, and Transaction Guarantees for Multi-Agent LLM Planning. Retrieved from https://dl.acm.org/doi/10.14778/3750601.3750611

[192] ArXiv. (2024, October 9). Constraint representation towards precise data-driven storytelling. Retrieved from https://arxiv.org/html/2410.07535v1

[195] Porteous, J., et al. Narrative Control using State Constraints. Retrieved from https://research.tees.ac.uk/ws/files/5954785/118625.pdf

[196] EmergentMind. (2025, November 26). Multi-Agent Reasoning Benchmark. Retrieved from https://www.emergentmind.com/topics/multi-agent-reasoning-driven-benchmark

[197] Dabral, C., et al. (2020). Constraint-Based Narrative Generation. In Proceedings of AIIDE. Retrieved from https://www.convivial.tools/PapersPublic/aiide2020-asp-narrative-gen.pdf

[200] Anthropic. (2025, June 12). How we built our multi-agent research system. Retrieved from https://www.anthropic.com/engineering/multi-agent-research-system