What changed
Oumi PBC today announced an update to its Frontier AI Engineer (FAiE) platform, enabling enterprises to build AI agents for customers and retrain the models those agents run on. This follows a common issue where agents fail after leaving the demo stage, often due to poor prompts, broken tools, or weak underlying models.
What this means in practice
The update allows teams to describe in plain language what they want an agent to accomplish, with FAiE helping define the tools and environments needed before moving to evaluation and training. If issues arise in production, they are turned into evaluation sets and training data for future improvements. Oumi also introduced direct preference optimization and reinforcement learning through group relative policy optimization, allowing models to be trained on user-preferred behaviors and task outcomes. These methods can be combined in a single training run, with FAiE assisting in strategy selection, reward setup, and problem diagnosis.
Limitations and availability
The new features, including group relative policy optimization, are launching in private preview, with Oumi accepting requests for access. The Bring Your Own Harness feature lets companies retain their orchestration, tools, and memory while connecting to Oumi for evaluation and model training. Oumi is a venture capital-backed startup that raised a $10 million seed round in January 2025. The supplied announcement does not specify the exact release timeline for all features.
