← Browse

World's Fair 2025

How to Train Your Agent: Building Reliable Agents with RL — Kyle Corbitt, OpenPipe

Kyle Corbitt

Overview

This talk details the process of building a reliable AI agent using reinforcement learning (RL), focusing on practical lessons learned from the ART E project, an email assistant. The core thesis is that while RL can significantly improve agent performance beyond prompted models, it's crucial to start with a strong prompted baseline, carefully design the training environment and reward functions, and be vigilant against reward hacking. The project demonstrates how RL can lead to substantial gains in accuracy, cost reduction, and latency improvements for specialized tasks.

Who should watch

Key takeaways

Notable quotes

*I would generally always recommend starting with getting the best performance you can with a prompted model before going to any training including reinforcement learning.*
*The last thing which we didn't apply here but can help a lot with these smaller things which is called speculative decoding.*
*It's really important to watch out for this solving it typically involves modifying in some way your reward function to penalize things like that.*

Watch on YouTube →

Unofficial community note. Prefer the recording for nuance.