Framework Core ALPHA
Experimental architectural patterns and building blocks derived from LLMunix's adaptive behavior management research on Gemini CLI. This is an ongoing exploration of reusable agent system components and design patterns for Gemini CLI-powered systems.
⚠️ Early-Stage Research - This project attempts to distill architectural insights from LLMunix research and will remain permanently in alpha status.
Key Innovations from LLMunix
🧠 Sentient State Patterns
Research into behavioral constraints that modify agent decision-making through dynamic constraint evolution, memory-driven learning, and modular state components with atomic file architectures.
🎯 Adaptive Behavior Research
Experimental investigations into user sentiment detection, priority adaptation mechanisms, persona switching prototypes, and error tolerance adjustment systems for dynamic behavior modification.
🔍 Memory System Experiments
Research prototypes investigating YAML frontmatter querying for execution history, behavioral pattern tracking, user preference evolution, and context-aware recommendation systems.
Architecture Patterns
1. Adaptive State Management Pattern
Based on LLMunix's sentient state architecture, this pattern provides a modular approach to managing agent execution state:
├── state/
│ ├── plan.md # Execution roadmap and metadata
│ ├── context.md # Accumulated knowledge base
│ ├── variables.json # Structured data passing
│ ├── history.md # Complete execution log
│ └── constraints.md # Behavioral modifiers (the sentient layer)
└── outputs/ # Task results and artifacts
2. Behavioral Constraint Pattern
Dynamic behavior modification through evolving constraints that adapt based on context and user interaction:
user_sentiment: "frustrated" # Detected emotional state
priority: "speed_and_clarity" # Execution focus
active_persona: "concise_assistant" # Communication style
error_tolerance: "strict" # Risk acceptance level
human_review_trigger_level: "low" # Escalation threshold
3. Memory-Driven Decision Pattern
Intelligent consultation of historical experiences to guide current decision-making:
---
timestamp: "2024-06-24T10:30:00Z"
task_type: "research_analysis"
user_sentiment_evolution: ["neutral", "impatient", "satisfied"]
constraint_adaptations:
- trigger: "user_time_pressure"
adaptation: "priority: comprehensiveness → speed_and_clarity"
success_metrics:
completion_rate: 0.95
user_satisfaction: 0.89
cost_efficiency: 0.76
lessons_learned:
- "Concise summaries preferred under time pressure"
- "Web scraping requires graceful degradation strategies"
---
How to Use This Framework
git clone https://github.com/EvolvingAgentsLabs/framework-core
2. Install Gemini CLI with virtual tools
git clone https://github.com/EvolvingAgentsLabs/gemini-cli
cd gemini-cli && git checkout issue-1806
# Follow build instructions
3. Create your "software"
# Add components/agents/ and components/tools/ directories
# Populate with markdown-defined agents and tools
4. Define your tasks
# Create scenarios/ directory with markdown task descriptions
# Create GEMINI.md manifest file for your system
5. Boot your OS
./llmunix-boot # Initialize the workspace
gemini # Start Gemini CLI runtime
# The GEMINI.md manifest will be automatically loaded
Future Vision & Roadmap
🌱 Autonomous Evolution & Self-Improvement
The system will analyze its own performance from SmartMemory.md to proactively evolve existing agents and tools for better efficiency, accuracy, and cost-effectiveness.
🛡️ Controlled Evolution (Firmware)
Introduction of Firmware.md component defining immutable rules, ethical constraints, and operational boundaries to guide agent evolution while maintaining safety protocols.
⚡ Parallel Task Processing
Development of a "Process Manager" agent and state management techniques allowing orchestration of multiple, concurrent task executions.
🔌 Multi-LLM Engine Support
Enhancement to be truly LLM-agnostic, dynamically choosing the best LLM for specific tasks and managing interactions across different AI engines.
🌐 Distributed Environments
Extension to distributed systems enabling multiple LLMUNIX instances to communicate, share components, and collaborate on complex goals.
🏗️ Building Blocks
Provides minimal "kernel" files that serve as the foundation for building your own LLM-OS instance with customizable components.
📚 Pattern Library
Distilled architectural patterns from LLMunix research including state management, behavioral constraints, and memory systems.
🔬 Research Foundation
Early-stage exploration of how adaptive behavior management concepts translate into generalizable architectural patterns.