Building Reliable Agents with RL – Kyle Corbitt, CEO of OpenPipe скачать в хорошем качестве

Building Reliable Agents with RL – Kyle Corbitt, CEO of OpenPipe

Why do AI agents still mess up the basics—and what can we do about it? In this talk, Kyle Corbitt breaks down how reinforcement learning (RL) can actually help us build agents that are way more reliable than just stacking prompts on prompts. He shares real-world examples of where agents go wrong, how to train them to behave better over time, and what it takes to define rewards that actually guide the right behavior. From debugging brittle agents to fine-tuning open-source models in the wild, Kyle walks through the nitty-gritty of making AI agents that don't fall apart when things get a little weird. Enterprise AI Agents Summit 2025 in Seattle. Hosted by OpenPipe + AWS on June 13, 2025.