Hugging Face publishes five-step guide to post-training AI agents

Hugging Face published “A curriculum for post-training agents” on Oct. 20, 2026. The community article sets out five steps for training AI agents after pretraining, with a coding agent as its running example. An agent can call tools, read their results and decide what to do next over multiple turns. The guide explains why post-training for agents must account for that whole sequence rather than a single answer.

The first step is supervised fine-tuning on agent traces, recordings of tool calls and their results usually produced by a stronger model. The guide notes a limitation: copying successful traces does not teach a model to recover from its own mistakes. Next comes distillation, in which a teacher model supervises a student through generated data or probabilities assigned to the student’s tokens.

The third step is reinforcement learning from scores on the model’s answers. The guide describes GRPO, a method that samples several answers to a prompt and compares their scores within the group. Step four adds an environment that keeps track of changes, runs tools and checks a final result across multiple turns, such as by running tests for a coding task. Step five trains on full agent runs through an existing harness inside remote sandboxes.