Oumi Expands Frontier AI Engineer to Let Customers Build and Retrain Their Own Agents

Oumi PBC, an enterprise artificial intelligence platform provider, has expanded its Frontier AI Engineer agent with capabilities to construct AI agents for customers and fine-tune the models those agents depend on. The enhancement tackles a recurring issue that typically emerges when agents transition from proof-of-concept to live deployment. When an agent underperforms, the root cause could stem from inadequate prompting, malfunctioning tools, or an insufficiently capable underlying model.

Historically, enterprises have relied on prompt engineering and tool refinement to address agent failures, since retraining model weights has traditionally demanded expertise concentrated at leading AI research organizations. This strategy frequently reaches a plateau before agents achieve the reliability needed for production use.

Oumi's Frontier AI Engineer already handled objective definition, evaluation design, and training experiment execution for proprietary models on its infrastructure. The latest update permits teams to articulate their agent requirements in natural language, with FAiE then assisting in specifying the tools and operational contexts the agent requires before progressing to evaluation and training phases. Production failures automatically convert into evaluation datasets and training material for subsequent iterations.

According to co-founder and Chief Executive Manos Koukoumidis, "by training the weights of the models powering them, learning from their own work," agents that achieve success will improve beyond what improved prompts and larger off-the-shelf models can deliver. He emphasized that failed tasks and user preferences represent the most valuable signals for this improvement.

Organizations operating agents on other platforms retain that flexibility. Through Oumi's Bring Your Own Harness feature, the customer's orchestration layer, tool integrations, and memory systems remain in their original location while connecting to Oumi's infrastructure for evaluation and model training. The resulting models become optimized specifically for the environment where the agent operates daily.

Oumi is introducing direct preference optimization alongside its existing supervised fine-tuning capability, permitting models to learn from examples of user-preferred behaviors. Reinforcement learning support arrives via group relative policy optimization, which assesses model performance based on task completion. Multiple training approaches can be combined in a single multistage training sequence, with FAiE managing strategy selection, reward configuration, and troubleshooting.

These training capabilities reinforce Oumi's central thesis regarding model ownership. Most enterprise agents currently operate on general-purpose models from major AI providers that lack optimization for specific customer use cases. Since launching its commercial platform for proprietary models in March, Oumi has argued that a specialized model owned entirely by the customer delivers superior performance at reduced expense.

Group relative policy optimization is entering private preview, with the Seattle-based company accepting access requests. Oumi, a venture-backed startup, secured a $10 million seed round in January 2025 led by Venrock Management and Obvious Management Services. The founding team comprises former engineering leaders from Google LLC's Gemini group.