Kyle Corbitt

Kyle Corbitt co-founded and led OpenPipe, which was acquired by CoreWeave. His AI work spans distilling expensive model workflows into smaller models and training multi-step agents with reinforcement learning. He is also a credited co-author of ART, the Agent Reinforcement Trainer, an open-source framework for agent training.[1][2][4]

Under Corbitt, OpenPipe initially focused on distilling expensive GPT-4 workflows into smaller, cheaper models, then pivoted to reinforcement learning for agents as frontier model prices fell, according to the description of his October 16, 2025 Latent Space episode. The shift moved the emphasis from reproducing an existing workflow to learning from an agent’s experience executing tasks.[1][4]

Corbitt shares ART’s author credit with Brad Hilton, David Corbitt, Saumya Gandhi, Angky William, Bohdan Kovalevskyi, and Andie Jones. ART provides a Python harness for training multi-step agents using GRPO, with a client that connects application code to a server handling inference and training.[4]

ART’s documented training loop records an agent’s messages in a trajectory and assigns a reward when the rollout finishes. The server groups trajectories for GRPO training, updates a LoRA adapter, and loads it into vLLM for the next round of inference. In this loop, inference pauses while training runs. Application code supplies the workflow and its reward; ART manages the model’s training cycle.[4]

RULER—Relative Universal LLM-Elicited Rewards— supplies another way to generate those rewards. Its documentation describes an LLM judge comparing several trajectories for the same scenario, deduplicating common prefixes, and assigning scores from zero to one based on goal achievement. The rationale is that comparing solutions is easier than scoring each in isolation, while GRPO normalizes scores within a group and therefore uses their relative differences.[5]

Corbitt emphasized the difficulty of sandboxing real systems with their bugs and edge cases intact and questioned GRPO’s long-term suitability because of its requirement for perfectly reproducible parallel rollouts. These concerns put the training environment alongside the learning algorithm as a practical obstacle to reliable agents.[1]

Corbitt also regarded LoRAs as underrated for production deployment and found GEPA and prompt optimization disappointing in his testing. His broader vision was for deployed agents to learn continuously from production experience, improving reliability through reinforcement learning. These technical concerns, production assessments, and ambitions are summarized in the October 2025 episode description.[1]

Before OpenPipe, Corbitt ran Y Combinator’s Startup School team and worked as a software engineer at Google. His earlier Startup School work provides a concrete example of his role building software and programs for founders.[2][3]

On June 17, 2020, Corbitt authored the announcement of Startup School’s relaunch as a continuous program, replacing the annual course’s fixed enrollment and graduation dates. The team focused on “curriculum, accountability and community,” informed by founder surveys and interviews. Changes included curriculum tailored to a company’s idea, launch, or growth stage; streaks for consecutive weekly updates; and improved matching for small-group video calls.[3]

Certification replaced graduation. Founders had to review the relevant curriculum, attend at least one weekly group session, and submit updates for eight consecutive weeks. The design let founders continue using Startup School after certification rather than treating the course as finished.[3]

The relaunch was a team effort. Corbitt credited product engineer Catheryn Li with reimagining core elements and building much of the new functionality. Joan DeGennaro joined as full-time community manager, focused on founder experience and inclusion. Corbitt’s author biography described his own work as creating software to support YC founders and noted that he had previously run a startup.[3]