Overview
Zilli is a next-generation Agent tool engineering framework for AI self-development. Its core philosophy is "AI writes AI", "evaluation as development", "from the environment", and "from Agent to RL". The entire system is built around a five-phase execution blueprint, forming a self-evolving closed loop.
Five-Phase Architecture
Phase 1: Definition
Define auto-verifiable task sets and type-safe Agent API contracts based on Pydantic.
Phase 2: Data
Build simulated sandbox environments, generate trajectory data, and establish a layered experience replay pool.
Phase 3: Infra
Deploy heterogeneous computing (SGLang + Megatron-LM) with adaptive length control and async rollout scheduling.
Phase 4: Model
CISPO algorithm for stable multi-turn Agent training, with RLVR reward shaping and GRPO baseline support.
Phase 5: Evolve
Offline evolution engine (DSPy + GEPA) optimizes Skills, with continuous learning from production data.
Relation to IClawOS
Zilli
AI Agent training infrastructure layer.
Provides RL training, Skill evolution, and trajectory data feedback for IClawOS agents.
IClawOS
AI-native Linux distribution, your 24/7 digital butler.
Zilli-trained models deploy to IClawOS, with production interaction data flowing back to Zilli for continuous improvement.
Tech Stack
Language
- Python 3.11+
Core Libraries
- pydantic
- numpy
- dspy-ai
RL Algorithms
- CISPO
- GRPO