FareedKhan-dev/training-ai-agents

Implements a multi-agent training pipeline using LangGraph and distributed RL algorithms (SFT, PPO, contextual bandits) with real-time observability via LangSmith and Weights & Biases. Agents collaborate through shared hierarchical state, exchange knowledge in parallel, and self-improve through dynamic reward systems that adapt based on performance and task alignment. The architecture progresses through supervised fine-tuning, reinforcement learning phases, and includes tracing hooks and logging adapters for capturing every interaction and learning step.