Building the Future of Intelligent Agents
Exploring adaptive AI systems through experimental frameworks and autonomous agent architectures
All projects are experimental research and will remain in permanent alpha status
Experiments
LLMunix 🦄 ALPHA
Pure Markdown Operating System
A revolutionary Pure Markdown Operating System designed to be run by multiple AI runtime engines. Compatible with Claude Code and Claude Code sub agents, Gemini CLI, and Qwen Code. Features multi-tier memory, inter-agent messaging, and dynamic evolution capabilities. Runtime engines interpret the manifest file to turn markdown specifications into a functional operating system.
Agent Forge ALPHA
JIT Agent Architecture Framework
Agent Forge is our comprehensive framework for just-in-time agent architectures, containing three specialized proof-of-concept implementations: jit-agent-poc (unified Qwen architecture), jit-agent-learn (reinforcement learning), and jit-agent-memory (persistent state). Features dynamic tool generation, benchmarkable performance improvements, and modular POC exploration.
├── jit-agent-poc ALPHA
Unified Qwen Architecture POC
Part of Agent Forge framework. Demonstrates a unified approach using a single fine-tuned Qwen2.5-Coder-1.5B model as both Orchestrator and Translator, eliminating multi-model complexity through specialized LoRA training.
├── jit-agent-learn ALPHA
Learning & Adaptation POC
Part of Agent Forge framework. Extension of the JIT agent architecture focused on reinforcement learning capabilities, allowing agents to improve their performance through experience and feedback loops using Qwen models.
└── jit-agent-memory ALPHA
Persistent Memory POC
Part of Agent Forge framework. The memory component of the JIT agent trilogy, adding persistent memory capabilities to enable contextual awareness and long-term information retention using Qwen models.
About Evolving Agents Labs
We're advancing the frontier of autonomous AI through experimental frameworks and research prototypes. Our work explores early-stage concepts in adaptive agent systems - all projects remain permanently in alpha status as ongoing research experiments.
📢 Our Research Evolution
Phase 1: Evolving Agents Toolkit (EAT) - Sunset
Our first project, the Evolving Agents Toolkit (EAT), was officially discontinued in July 2025. While EAT demonstrated powerful concepts in multi-agent orchestration with MongoDB backend, we recognized that the complex Python architecture was over-engineered for achieving adaptive agent behavior.
Phase 2: LLMunix - Simplified Evolution
EAT's concepts were dramatically simplified and reimplemented in LLMunix - a Pure Markdown Operating System that achieves the same adaptive agent goals through elegant simplicity. From EAT's multi-component Python architecture with MongoDB backend to LLMunix's pure markdown definitions interpreted by LLM runtime engines - same adaptive capabilities, 10x simpler implementation.
Phase 3: Agent Forge JIT POCs - Current Focus
With Claude Code's implementation of sub-agents in markdown as an official feature, we realized our original markdown-based agent concept was validated. We now focus on Agent Forge and JIT POCs - exploring just-in-time compilation, hybrid architectures, and benchmarkable performance improvements over pure LLM approaches.
Adaptive Behavior Research
Experimental systems that explore how agents might modify their decision-making processes based on context and interaction patterns.
Pure Markdown Architecture
Exploring the use of markdown as a full operating system specification, enabling clean separation of behavior, state, and execution logic.