Oumi's Frontier AI Engineer now builds agents and retrains their models

Oumi says many enterprise agents stall because prompt and harness fixes cannot repair a weak underlying model, and retraining model weights has typically required skills found at frontier labs. With the update, a team describes its goal in plain language, and the Frontier AI Engineer helps define the tools and environments the agent needs before evaluation and training. Failures seen in production are turned into evaluation sets and training data for the next round.

Customers already running agents elsewhere can keep them there under Bring Your Own Harness, where orchestration, tools and memory stay in the customer's environment and connect to Oumi for evaluation and training. On the training side, Oumi is adding direct preference optimization alongside its existing supervised fine-tuning, plus group relative policy optimization for reinforcement learning, which grades a model on task outcomes. Several methods can be chained together in one multistage training run.

Developers can tune the model behind an agent in the same environment where the agent runs, rather than relying only on prompt and tool changes. Reinforcement learning through group relative policy optimization is in private preview, so access requires requesting it from Oumi. Teams taking this route would need to plan for evaluation sets and training data built from production failures.