16 KiB
16 KiB
AgentDB Learning: Visual Guide
Purpose: Visual diagrams and flow charts showing exactly how AgentDB learns and improves skill creation.
🔄 The Complete Learning Loop (Visual)
Macro Level: Creation → Learning → Improvement
┌─────────────────┐ ┌──────────────────┐ ┌─────────────────┐
│ User Request │───▶│ Agent Creator │───▶│ Skill Created │
│ │ │ │ │ │
│ "Create agent │ │ Uses: │ │ Functional code │
│ for stocks" │ │ • /references │ │ • Documentation │
└─────────────────┘ │ • AgentDB data │ │ • Tests │
└──────────────────┘ └─────────────────┘
│ │
▼ ▼
┌──────────────────┐ ┌─────────────────┐
│ Store in AgentDB│───▶│ Deploy Skill │
│ │ │ │
│ • Episodes │ • User starts │
│ • Causal edges │ • using skill │
│ • Success data │ • Provides feedback│
└──────────────────┘ └─────────────────┘
│ │
▼ ▼
┌─────────────────┐ ┌──────────────────┐ ┌─────────────────┐
│ Future User │◀───│ AgentDB Query │◀───│ Learning Data │
│ Request │ │ │ │ Accumulated │
│ │ • Similar past │ │ │
│ "Create agent │ • Success rates │ • Better patterns│
│ for crypto" │ • Proven templates │ • Higher success │
└─────────────────┘ └──────────────────┘ └─────────────────┘
📊 Data Storage Structure (Visual)
What Gets Stored Where in AgentDB
AgentDB Database
├── 📚 Episodes (Reflexion Store)
│ ├── Episode #1
│ │ ├── session_id: "creation-20251024-103406"
│ │ ├── task: "agent_creation_decision"
│ │ ├── input: "Create financial analysis agent..."
│ │ ├── reward: 85.0
│ │ ├── success: true
│ │ └── template_used: "financial-analysis-template"
│ │
│ ├── Episode #2
│ │ ├── session_id: "creation-20251024-103456"
│ │ ├── task: "agent_creation_decision"
│ │ ├── input: "Build climate analysis tool..."
│ │ ├── reward: 0.0
│ │ ├── success: false
│ │ └── template_used: "climate-analysis-template"
│ │
│ └── ... (one episode per creation)
│
├── 🔗 Causal Edges
│ ├── Edge #1
│ │ ├── cause: "finance_domain_request"
│ │ ├── effect: "financial_template_selected"
│ │ ├── uplift: 0.25
│ │ ├── confidence: 0.85
│ │ └── sample_size: 12
│ │
│ ├── Edge #2
│ │ ├── cause: "climate_domain_request"
│ │ ├── effect: "climate_template_selected"
│ │ ├── uplift: 0.30
│ │ ├── confidence: 0.90
│ │ └── sample_size: 8
│ │
│ └── ... (learned cause→effect relationships)
│
└── 🛠️ Skills Database
├── Skill #1
│ ├── name: "financial-pattern-skill"
│ ├── description: "Common patterns for finance agents"
│ ├── success_rate: 0.82
│ ├── uses: 15
│ └── learned_features: ["RSI", "MACD", "volume"]
│
└── ... (extracted patterns from successful episodes)
🔍 Query Process (Step-by-Step Visual)
When User Requests: "Create financial analysis agent"
Step 1: Input Analysis
┌─────────────────────────────────────┐
│ User Input: "Create financial │
│ analysis agent for stocks" │
│ │
│ → Extract domain: "finance" │
│ → Extract features: "analysis", │
│ "stocks" │
│ → Generate search queries │
└─────────────────────────────────────┘
│
▼
Step 2: AgentDB Queries
┌─────────────────────────────────────┐
│ Query 1: Episodes │
│ agentdb reflexion retrieve │
│ "financial analysis" 5 0.6 │
│ │
│ Query 2: Causal Effects │
│ agentdb causal query │
│ "use_finance_template" "" 0.7 │
│ │
│ Query 3: Skills Search │
│ agentdb skill search │
│ "financial analysis" 5 │
└─────────────────────────────────────┘
│
▼
Step 3: Data Analysis
┌─────────────────────────────────────┐
│ Episodes Retrieved: │
│ ┌─ Episode A: Success=True │
│ │ Template: financial-template │
│ │ Reward: 85.0 │
│ └─ Episode B: Success=False │
│ Template: generic-template │
│ Reward: 0.0 │
│ │
│ Success Rate: 50% (1/2) │
│ │
│ Causal Effects Found: │
│ ┌─ financial-template: uplift=0.25 │
│ └─ generic-template: uplift=0.10 │
└─────────────────────────────────────┘
│
▼
Step 4: Decision Making
┌─────────────────────────────────────┐
│ Decision Factors: │
│ ✓ 25% uplift for financial-template │
│ ✓ 50% historical success rate │
│ ✓ Domain match: "finance" │
│ │
│ Enhanced Decision: │
│ → Template: financial-template │
│ → Confidence: 0.50 │
│ → Proof: "Causal uplift: 25%" │
│ → Features: ["RSI", "MACD"] │
└─────────────────────────────────────┘
📈 Learning Progression (Visual Timeline)
How the System Gets Smarter Over Time
Month 1: Initial Learning
┌─────────────────────────────────────┐
│ Creations: 5 │
│ Episodes: 5 │
│ Success Rate: Unknown │
│ Templates: Static from /references │
│ Learning: Basic pattern recording │
└─────────────────────────────────────┘
Month 3: Pattern Recognition
┌─────────────────────────────────────┐
│ Creations: 25 │
│ Episodes: 25 │
│ Success Rates: Emerging │
│ Templates: Domain-specific patterns │
│ Learning: Success rate calculation │
└─────────────────────────────────────┘
Month 6: Intelligent Recommendations
┌─────────────────────────────────────┐
│ Creations: 100 │
│ Episodes: 100 │
│ Success Rates: Reliable (>10 samples)│
│ Templates: Optimized per domain │
│ Learning: Causal relationship mapping│
└─────────────────────────────────────┘
Month 12: Expert System
┌─────────────────────────────────────┐
│ Creations: 500+ │
│ Episodes: 500+ │
│ Success Rates: Highly accurate │
│ Templates: Self-optimizing │
│ Learning: Predictive recommendations │
└─────────────────────────────────────┘
🎯 Real Example: From First to Tenth Creation
Creation #1: No Learning Data
User: "Create financial analysis agent"
Process:
┌─ Query episodes: 0 results
├─ Query causal: 0 results
├─ Query skills: 0 results
└─ Decision: Use /references guidelines
Result:
┌─ Template: financial-analysis (from /references)
├─ Confidence: 0.8 (base rate)
├─ Features: Standard set
└─ Storage: Episode + Causal edge recorded
Creation #10: Rich Learning Data
User: "Create financial analysis agent for crypto"
Process:
┌─ Query episodes: 8 similar results
│ ├─ Success: 6/8 = 75% success rate
│ └─ Common features: ["RSI", "volume", "volatility"]
│
├─ Query causal: 5 relevant edges
│ ├─ financial-template: uplift=0.25
│ ├─ crypto-specific: uplift=0.15
│ └─ volatility-analysis: uplift=0.10
│
└─ Query skills: 3 relevant skills
├─ crypto-analysis-skill: success_rate=0.82
├─ technical-indicators-skill: success_rate=0.78
└─ market-data-skill: success_rate=0.85
Result:
┌─ Template: financial-analysis-enhanced
├─ Confidence: 0.75 (from historical data)
├─ Features: ["RSI", "MACD", "volatility", "crypto-specific"]
├─ Proof: "Causal uplift: 25% + crypto patterns: 15%"
└─ Storage: New episode + refined causal edges
🔧 Technical Flow Diagram
Code-Level Data Flow
enhance_agent_creation(user_input, domain)
│
▼
┌─────────────────────────────────────────┐
│ Step 1: Query Historical Episodes │
│ episodes = query_similar_episodes(input)│
│ │
│ SQL equivalent: │
│ SELECT * FROM episodes │
│ WHERE similarity(input, task) > 0.6 │
│ ORDER BY similarity DESC │
│ LIMIT 3 │
└─────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────┐
│ Step 2: Calculate Success Patterns │
│ success_rate = successful/total │
│ │
│ if success_rate > 0.7: │
│ prefer_this_pattern = True │
└─────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────┐
│ Step 3: Query Causal Relationships │
│ effects = query_causal_effects(domain) │
│ │
│ SQL equivalent: │
│ SELECT * FROM causal_edges │
│ WHERE cause LIKE '%domain%' │
│ AND uplift > 0.1 │
│ ORDER BY uplift DESC │
└─────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────┐
│ Step 4: Search Learned Skills │
│ skills = search_relevant_skills(input) │
│ │
│ SQL equivalent: │
│ SELECT * FROM skills │
│ WHERE similarity(description, query) > 0.7│
│ AND success_rate > 0.6 │
└─────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────┐
│ Step 5: Make Enhanced Decision │
│ intelligence = AgentDBIntelligence( │
│ template_choice=best_template, │
│ success_probability=success_rate, │
│ learned_improvements=extract_features(skills),│
│ mathematical_proof=causal_proof │
│ ) │
└─────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────┐
│ Step 6: Store for Future Learning │
│ store_creation_decision(input, intelligence)│
│ │
│ SQL equivalent: │
│ INSERT INTO episodes VALUES (...) │
│ INSERT INTO causal_edges VALUES (...) │
└─────────────────────────────────────────┘
🎉 Key Takeaways (Visual Summary)
┌─────────────────────────────────────────┐
│ AgentDB Learning Magic │
│ │
│ 📚 Store Every Decision │
│ 🔍 Find Similar Past Decisions │
│ 📊 Calculate Success Patterns │
│ 🎯 Make Enhanced Recommendations │
│ 🔄 Continuously Improve │
│ │
│ Result: System gets smarter with │
│ every skill created! │
└─────────────────────────────────────────┘
From "nebulous magic" to "understandable process" - AgentDB turns Agent Creator into a learning system that accumulates expertise with every interaction!