⭐ If you find this repository helpful, please consider giving it a ⭐ here on GitHub (click the star button in the top right corner) It's a quick way to show support for this openly available code. ⭐

Develop Self-Improving AI Agents with Reinforcement Learning

This repository provides the hands-on excercises for the Live Event. It covers policy rollouts, reward modeling, trajectory generation, optimization methods, and tool-use training for agentic systems.

Repository Structure

.
├── hands_on/  # Hands-on exercises
├── helper_functions/  # Helper functions for some notebooks     
└── README.md

---

---

Contents and Exercises

Section 1 • Foundations of RL for LLMs

Core concepts:

• Policies, rollouts, trajectories, rewards
• Minimal reinforcement loop for LLMs
• Relative feedback integration

Hands on notebook:

---

Section 2 • Reward Modeling and Relative Scoring

Core concepts:

• Reward shaping for LLM behavior
• Relative ranking vs absolute scoring
• RULER relative universal LLM elicited rewards

Hands on notebook:

---

Section 3 • Agent Trajectories and Rollout Design

Core concepts:

• Agent Reinforcement Trainer rollout generation
• Message sequences, actions, and feedback
• Multi trajectory sampling for optimization

Hands on notebook:

---

Section 4 • Optimization and Continuous Improvement

Core concepts:

• GRPO group relative policy optimization
• GSPO group sequence policy optimization
• Stabilizing RL for large models including MoE architectures

Hands on notebook:

---

Section 5 • Teaching Agents Tool Mastery (MCP Integration)

Core concepts:

• Automatic tool use training via structured task sets
• Using RULER for tool use reward evaluation
• MCP based agent to agent communication

Hands on notebook:

---

Running the Notebooks

All notebooks are designed to run in Google Colab without local setup.
Click the corresponding Open in Colab button above to start an exercise.