In this OpenAI Build Hour, engineers describe AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → RFT, a form of reinforcement fine-tuningTaking a trained model and training it a bit more on your own examples so it gets better at one specific job.Full definition →, where a model is trained by trial and scored reward. OpenAI says the base version could not fine-tune agents. Agent RFT adds two things: the model can call tool endpoints you host while it explores, and you supply an endpoint that returns a custom reward score.
Each rollout gets a unique ID attached to its tool calls so your system can tie them to the final answer and grade with full context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition →. They say this trains models for your real tools and can reduce tool calls or keep them within a budget, which lowers latency. They also say it is sample efficient. These are OpenAI's design claims, and the excerpt does not show results.
Terms in this piece · Glossary
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
fine-tuning — Taking a trained model and training it a bit more on your own examples so it gets better at one specific job.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.