What is agentic-rl-from-scratch?

This is a teaching repository that walks through how AI agents learn through reinforcement, using only Python's built-in standard library for the core examples. There is no GPU required, no calls to any AI service, and no machine-learning framework to install for the first three notebooks. The goal is to make the concepts behind Agentic RL something you can read, run, and trace line by line on a laptop in a few seconds. The repository is aimed at people who already build AI agent workflows but find the vocabulary around reinforcement learning hard to separate. Terms like RLHF, DPO, GRPO, and RLVR often blur together. The opening docs chapter distinguishes them before any code runs. The notebooks then build up the training loop piece by piece: a toy environment that teaches how rewards turn into policy updates, then a multi-turn math agent that introduces trajectories and loss masks, then a full GRPO training loop that takes an agent from using tools randomly to using them correctly. Every notebook ships with its outputs already included, and the random seed is fixed, so each intermediate result, such as how reward breaks down per turn or how the loss mask selects which tokens count, is printed and reproducible without running anything. You can open them on GitHub and read the full walkthrough without installing a single package. Two optional notebooks bridge the teaching concepts to production frameworks. One shows how the reward and advantage structures from the earlier chapters map onto the GRPOTrainer from a library called TRL. The other does the same for verl. These optional chapters require PyTorch and TRL, which the repository keeps in a separate requirements file precisely to keep the core examples dependency-free. The repository explicitly states what it does not do: it does not reimplement production-grade training algorithms and is not a replacement for TRL, verl, or OpenRLHF. Its value is clarity over scale.