Agent Reinforcement Fine-Tuning (Agent RFT) is OpenAI's method for improving AI agents that interact with external tools by changing model weights based on custom reward signals. During training, agents explore different tool-calling strategies while the model learns from rollouts that interact with real endpoints. The technique is sample-efficient (working with as few as 10 examples), reduces latency by optimizing tool call budgets, and helps models adapt to domain-specific environments. Case studies from Cognition, Codto, Cosine, and Macco demonstrate 5-10% accuracy improvements and significant latency reductions. Success requires well-defined tasks, production-matching training data, explorable solution spaces, and robust reward functions that resist gaming.
Sort: