📖 Agent Categories

Legend: 👍 Ready now · 🤔 Future development · 👎 Long-term · ⛔ Avoiding

🎨 Design Patterns

Structural Patterns

• Assembly Line - Sequential specialized pipeline for repeatable workflows
• Modular Cognitive - Specialized interacting modules for complex reasoning
• Multi-Agent Collaboration - Team of specialists cooperating toward common goals

Interaction Patterns

• Complementary Pair - Balanced dual perspectives (e.g., tech + UX)
• Adversarial Debate - Truth-seeking through competitive scrutiny
• Epistemic Delegation - Task delegation to specialized components

Process Patterns

• Self-Reflective Loop - Iterative self-critique and refinement
• Active Memory Management - Context curation and intelligent persistence

See full pattern documentation →

🔬 Research Highlights

Real-World Results

Key Insights

"We must move beyond one-shot prompt→response paradigms and equip agents with internal structure and interactions that mirror the complexities of reasoning."

The research converges on a clear trend: architectural modularity, feedback-rich loops, and role abstraction grounded in computational functions rather than human job titles.

🌊 Spec-Driven Development (SDD)

Spiral Grove implements a rigorous four-phase methodology:

Phase 1: Specification

Define WHAT to build

• User stories and success criteria
• Functional/non-functional requirements
• Explicit constraints

Phase 2: Planning

Design HOW to build it

• Technical architecture
• Decision rationale
• Integration points

Phase 3: Task Breakdown

Create STEPS to execute

• Discrete, independent tasks
• Dependencies and ordering
• Acceptance criteria

Phase 4: Implementation

Execute and TRACK progress

• One task at a time
• Continuous validation
• Deviation documentation

Learn more about SDD →

🗂️ Repository Structure

vibe-garden/
├── seeds/                       # Research and brainstorming
│   ├── brainstorm/
│   │   ├── agents/             # 30+ agent types documented
│   │   └── patterns/           # 8 major design patterns
│   ├── notes/                  # Academic research and theory
│   └── scripts/                # Utility scripts
├── spiral-grove/               # SDD Claude Code plugin
│   ├── .claude-plugin/         # Plugin configuration
│   └── commands/               # SDD phase implementations
└── .sdd/                       # Active SDD artifacts
    ├── specs/                  # Feature specifications
    ├── plans/                  # Technical plans
    ├── tasks/                  # Task breakdowns
    └── progress/               # Implementation tracking

🧩 Core Principles

Agent Design Philosophy

1. Functional Abstraction - Define agents by computational functions, not job titles
2. Modular Architectures - Specialized, interacting components over monolithic systems
3. Feedback Loops - Enable self-reflection and iterative improvement
4. Epistemic Delegation - Leverage specialists for knowledge-intensive tasks
5. Deviation Detection - Dedicated error monitoring and quality assurance

SDD Methodology

1. Specs Define Success - Specification is source of truth
2. Document Trade-offs - Always explain WHY, not just WHAT
3. Stay in Phase - Complete one phase before moving to next
4. Validate Continuously - Map implementation to spec acceptance criteria
5. Track Deviations - Document when/why implementation differs

🚀 Use Cases

For AI Researchers

• Comprehensive taxonomy of agent types and patterns
• Academic references and performance metrics
• Emerging theoretical models and frameworks

For Developers

• Structured development methodology (SDD)
• Reusable agent design patterns
• Claude Code plugin for systematic feature development

For Teams

• Shared vocabulary for agent architectures
• Decision frameworks for agent selection
• Documented rationale for technical choices

📊 Pattern Selection Guide

By Project Type

Software Development: Assembly Line + Complementary Pair + Self-Reflective

Research & Analysis: Multi-Agent + Epistemic Delegation + Active Memory

Decision Making: Modular Cognitive + Adversarial Debate + Self-Reflective

User-Facing Products: Complementary Pair + Active Memory + Self-Reflective

By Challenge

• Hallucinations/Errors → Epistemic Delegation, Adversarial Debate
• Complex Multi-Step Tasks → Modular Cognitive, Assembly Line
• User Experience → Complementary Pair with Empathy Agent
• Learning from Mistakes → Self-Reflective Loop, Active Memory
• Long Context → Active Memory Management
• Coordination → Multi-Agent with Manager/Orchestrator

🛠️ Utilities

PDF Converter

Convert research papers to markdown:

python seeds/scripts/convert_pdf.py <pdf_file> [output_file]

Preserves layout and formatting for integration with Claude Code.

📚 Documentation

• Agent Types Index - Complete agent taxonomy
• Design Patterns Guide - All 8 patterns
• Emerging Theory - Theoretical foundations
• SDD Workflow - Spec-Driven Development guide
• Claude Code Integration - Guidance for Claude Code instances

🌟 Evolution & Future Directions

Trajectory

Monolithic LLM
    ↓
Single LLM with tools
    ↓
Pipeline of specialized agents
    ↓
Modular cognitive architecture
    ↓
Multi-agent teams with diverse roles
    ↓
Self-improving agent societies  ← We are here (2024-2025)

Emerging Trends

• Learned coordination strategies
• Dynamic team formation
• Hierarchical agent systems
• Agent societies and ecosystems
• Seamless human-AI collaboration
• Formal verification of agent behavior

🤝 Contributing

This is a living research repository. Contributions welcome:

1. New Agent Types - Document emerging agent architectures
2. Design Patterns - Add proven patterns with examples
3. Research Notes - Include academic papers and findings
4. SDD Improvements - Enhance the Spiral Grove plugin
5. Real-World Examples - Share implementation experiences

Please follow the established documentation format and include academic citations where applicable.

📄 License

MIT License - see LICENSE file for details.

🔗 Resources

Academic Papers

All patterns and agent types reference academic research from 2018-2025. See individual documentation files for detailed citations.

Related Projects

• Claude Code - Official Anthropic CLI

📮 Contact

Author: Ronald Roy Email: gsdwig@gmail.com Repository: github.com/rjroy/vibe-garden

Built with 🧠 for the future of AI agent design

Last Updated: 2025-10-16