AgentTuning fine-tunes LLMs on interaction trajectories to imbue them with agent abilities such as multi-step reasoning, tool use, and task execution. The project provides the AgentInstruct dataset containing 1,866 curated interactions spanning 6 real-world task scenarios, along with the resulting AgentLM models (up to 70B parameters) released on HuggingFace. The approach uses ReAct-style reasoning traces and rigorous quality filtering to ensure both agent generalization and retained general language capabilities.