"AlphaGo Moment for Model Architecture Discovery" - Implementing evolutionary AI for autonomous project architecture generation
ASI-ARCH is an advanced framework for Claude Code that implements autonomous architecture discovery through evolutionary AI principles. Based on the groundbreaking research paper that discovered 106 novel architectures through 1,773 experiments, this framework adapts the methodology using a 3-module system (Pipeline, Database, Cognition Base) to automatically discover optimal project architectures.
- 🧬 Evolutionary Architecture Discovery: Continuously evolves architectures through multiple generations
- 🤖 22 Specialized Agents: Including LLM Judge and Deduplicator for rigorous evaluation
- 📊 Composite Fitness Function: Performance (70%) + LLM Judge Score (30%) with sigmoid amplification
- 💡 Two-Stage Strategy: Fast exploration followed by deep verification
- 📈 Empirical Scaling Law: Tracks how fitness improves with computational investment
- 🔄 Deduplication System: Ensures genuine innovation by preventing redundant exploration
- 📊 Experiment Tracking: Target of 1,773 experiments matching the original research
┌─────────────────┐ ┌─────────────────────────────────────────┐
│ Cognition Base │────▶│ Pipeline Module │
│ (Knowledge) │ │ ┌──────────┐ ┌─────────┐ ┌────────┐ │
└─────────────────┘ │ │Researcher│──│Engineer │──│Analyst │ │
▲ │ └──────────┘ └─────────┘ └────────┘ │
│ └─────────────────────────────────────────┘
│ │
│ ┌───────────────▼───────────────┐
└──────────────│ Database Module │
│ (Experiments & Candidates) │
└──────────────────────────────┘
| Module | Key Agents | Purpose |
|---|---|---|
| Pipeline | Researcher (hypothesis-generator, cross-pollinator), Engineer (code-architect, quality-guardian), Analyst (insight-miner, llm-judge) | Autonomous discovery loop |
| Database | deduplicator, memory-architect, experiment tracker | Store architectures and prevent redundancy |
| Cognition Base | pattern-miner, knowledge-synthesizer | Curated knowledge repository |
| Support | visualizer, orchestrator, meta-learner | System coordination and improvement |
/asi-arch Create a real-time collaborative code editor with AI assistance# Multiple evolution cycles for complex problems
/asi-arch --evolve Design a distributed ML training platform
# Interactive evolution with user feedback
/asi-arch --guide Build a blockchain-based supply chain system
# Export successful architecture as template
/asi-arch --export-template microservices-auth Design authentication microservice.claude/
├── slash-commands/
│ └── asi-arch.md # Main slash command
├── agents/
│ ├── asi-arch-*.md # 20 specialized agents
│ └── ...
├── asi-arch-context.md # Context management rules
└── asi-arch/ # Runtime data
├── knowledge-base/ # Discovered patterns
├── evolution-history/ # Generation tracking
└── active-session/ # Current execution
- Clone this repository:
git clone https://github.com/Swarek/discovery_agents.git
cd discovery_agents- The framework is automatically available in Claude Code once the repository is opened.
Generation 1: Two-phase commit protocol (Fitness: 72)
Generation 3: Saga pattern with compensation (Fitness: 85)
Generation 5: Event sourcing with CQRS (Fitness: 94)
Generation 1: Batch processing pipeline (Fitness: 68)
Generation 4: Stream processing with Lambda architecture (Fitness: 87)
Generation 7: Hybrid with ML-driven optimization (Fitness: 96)
- Context Extraction: Analyzes your project requirements and constraints
- Pattern Discovery: Mines successful architectures from the knowledge base
- Hypothesis Generation: Creates novel architectural candidates using evolutionary algorithms
- Deduplication Check: Ensures genuine novelty (>85% different from existing)
- Two-Stage Evaluation:
- Exploration: Quick evaluation on simplified problems
- Verification: Deep validation of top candidates
- LLM Judge: Evaluates novelty and elegance (30% of score)
- Evolution: Continues until convergence or experiment limit
- Scaling Law: Tracks log(fitness) vs log(experiments) relationship
Each architecture is evaluated on:
Composite Fitness = σ(Performance) × 0.7 + LLM_Judge_Score × 0.3
Where:
- σ(x) = 1 / (1 + e^(-10(x-0.5))) - Sigmoid to amplify meaningful gains
- Performance: Objective metrics (speed, scalability, efficiency)
- LLM Judge Score: Novelty, elegance, breakthrough potential, practical impact
ASI-ARCH is designed to evolve and improve. Contributions are welcome!
- Use the framework and share your discovered architectures
- Report issues or suggest improvements
- Contribute new cross-domain patterns
- Share success stories and case studies
This implementation is inspired by the paper "AlphaGo Moment for Model Architecture Discovery" which demonstrated autonomous AI research capabilities in discovering novel linear attention architectures.
- Architecture Marketplace: Share and discover community patterns
- Domain Specialization: Pre-trained models for specific industries
- Architecture Pattern Analyzer: AI-driven pattern discovery engine
- Performance Predictions: ML-based architecture performance forecasting
MIT License - See LICENSE file for details
- Original ASI-ARCH research team
- Claude Code community
- Contributors and early adopters
"The best architecture is not designed, it's discovered through evolution."
Ready to discover your next breakthrough architecture? Try /asi-arch today!