RL Dojo vs Harness Dojo: Two Distinct Approaches to Training AI Agents Explained

A technical explainer published on September 26, 2026, outlines two fundamentally different methods used to train AI agents, collectively referred to as 'Dojo' approaches. The first, RL Dojo, uses reinforcement learning to directly update a model's weights based on rewards earned during training, as demonstrated by Meta's Kaggle-winning agent and the MLE-Dojo project from Georgia Tech and Stanford. The second, Harness Dojo, leaves model weights unchanged and instead improves agent behavior by refining surrounding elements such as instructions, rules, memory, and stored lessons — illustrated by the open-source copilot-agents-dojo project on GitHub. RL Dojo produces durable, environment-independent skills but demands significant compute resources, labeled data, and machine learning expertise, while Harness Dojo is low-cost and transparent but remains bounded by the base model's capabilities. Many teams ultimately combine both approaches, using Harness Dojo for rapid iteration and RL Dojo when resources allow deeper model-level improvements.