Summary

OpenAI researchers discuss agent development, focusing on how artificial intelligence agents can interact with external tools to complete complex tasks autonomously. The presentation introduces Codeex, a flagship coding agent, and explores strategies for improving agent performance including prompt engineering, task optimization, and fine-tuning techniques. The key takeaway is agent reinforcement fine-tuning (agent RF), which allows developers to modify model weights by specifying learning signals that define desired and undesired agent behaviors, ultimately enhancing an agent's performance on specific tasks.