Oumi’s AI Engineer adds agent building and model retraining capabilities

Oumi PBC stated that its Frontier AI Engineer agent now builds AI agents for customers and retrains the models those agents run on, according to a report by SiliconANGLE AI.reported

The claim

SiliconANGLE AI reported that Oumi extended its platform to address failures that emerge when agents move past demos.reported The company said failing agents often suffer from bad prompts, broken tools or weak underlying models. Traditionally, enterprises attempted fixes through prompt and harness engineering because retraining model weights required skills found at frontier AI labs. Oumi claims this approach often stalls before agents become reliable enough for production work.

The response

Co-founder and Chief Executive Manos Koukoumidis said better prompts and bigger generic models have limits. He stated that successful agents improve by training the weights of the models powering them and learning from their own work. Koukoumidis pointed to failed tasks and user preferences as the signals that count for this improvement process.

What the evidence means

The material indicates that Oumi is positioning model weight training as a solution to agent reliability issues. The company argues that most enterprise agents run on general-purpose models from frontier providers that are not optimized for specific customer work. This suggests a shift toward custom-tuned models for enterprise use cases.

In practice

Teams can describe in plain language what they want an agent to accomplish. The Frontier AI Engineer helps define the tools and environments the agent needs before moving to evaluation and training. Production failures are turned into evaluation sets and training data for the next round. Under Bring Your Own Harness, customers keep their orchestration, tools and memory while connecting to Oumi for evaluation and model training. Models are tuned in the same environment the agent works in daily. The platform adds direct preference optimization to supervised fine-tuning and reinforcement learning through group relative policy optimization. Multiple methods can be chained in a single multistage training run.