Oumi Extends Frontier AI Engineer Capabilities for Enhanced AI Agent Development
Oumi PBC, an enterprise artificial intelligence platform company, has announced an expansion of its Frontier AI Engineer (FAiE) agent, enabling it to construct AI agents for customers and retrain the models underlying those agents. This development aims to address challenges that arise when agents transition from the demonstration phase to operational deployment.
Oumi identifies a common issue where agents may falter due to poorly designed prompts, malfunctioning tools, or suboptimal underlying models. Traditional remedies in enterprises often involve prompt and harness engineering, a process that typically requires expertise associated with cutting-edge AI labs. This approach can hinder the reliability necessary for agents to operate effectively in the real world.
The FAiE system is designed to set objectives, create evaluations, and conduct training experiments for custom models directly on the platform. With the latest update, teams can articulate their goals in straightforward language, and FAiE will assist in determining the necessary tools and environments before proceeding to evaluation and training. Any issues that arise during production can be converted into evaluation datasets and training data for future iterations.
Manos Koukoumidis, Co-founder and Chief Executive of Oumi, noted that simply improving prompts and utilizing larger generic models won’t suffice for long-term success. According to him, the agents that truly excel will evolve “by training the weights of the models powering them, learning from their own work,” utilizing feedback from unsuccessful tasks and user preferences as critical indicators.
Oumi also supports customers who manage AI agents in other environments with its “Bring Your Own Harness” feature, allowing them to retain their orchestration, tools, and memory while seamlessly connecting to Oumi for evaluation and model training. This means that models can be optimized within the same operational context in which the agents function daily.
Additionally, Oumi is integrating direct preference optimization alongside its existing supervised fine-tuning capabilities, allowing for models to be trained using examples of user-preferred behaviors. The introduction of reinforcement learning through group relative policy optimization will evaluate model performance based on task outcomes. Multiple methodologies can be interlinked in a single multistage training process, with FAiE facilitating strategy selection, reward setup, and problem diagnosis throughout the training cycle.
This advancement aligns with Oumi’s broader mission regarding model ownership. Many enterprise agents currently rely on general-purpose models provided by leading AI firms that are often not tailored to specific customer needs. Since launching its commercial platform for custom models in March, Oumi has been advocating for specialized models that clients fully control, asserting they can achieve superior performance at a reduced cost.
Group relative policy optimization is currently in a private preview phase, and the Seattle-based company is accepting requests for access. Oumi secured a $10 million seed investment round in January 2025, led by Venrock Management and Obvious Management Services, and boasts a founding team of former engineering leaders from Google LLC’s Gemini group.