Oumi 扩展 Frontier AI Engineer：可构建 AI 智能体并重训其底层模型

Enterprise artificial intelligence platform company Oumi PBC today extended its Frontier AI Engineer agent so it can build AI agents for customers and retrain the models those agents run on.

The release goes after a problem that tends to surface once agents leave the demo stage. A failing agent might have a bad prompt, a broken tool or a weak underlying model.

Oumi says enterprises have traditionally tried to fix it with prompt and harness engineering because retraining a model’s weights has required skills normally found at frontier AI labs. That approach often stalls before the agent is reliable enough to put to work.

Oumi’s Frontier AI Engineer already sets objectives, designs evaluations and runs training experiments for custom models on its platform. With today’s update, a team can describe in plain language what it wants an agent to accomplish, and FAiE helps define the tools and environments that agent needs before moving on to evaluation and training. Whatever goes wrong once the agent is in production is turned into evaluation sets and training data for the next round.

Co-founder and Chief Executive Manos Koukoumidis said better prompts and bigger generic models will only take agents so far. The ones that succeed will improve “by training the weights of the models powering them, learning from their own work,” he said, pointing to failed tasks and user preferences as the signals that count.

Companies that already run agents elsewhere can keep them there. Under a feature Oumi calls Bring Your Own Harness, the customer’s orchestration, tools and memory stay put and connect to Oumi for evaluation and model training. Models end up tuned in the same environment the agent works in every day.

Oumi is also adding direct preference optimization to the supervised fine-tuning it already supported, so a model can be trained on examples of the behavior users preferred. Reinforcement learning arrives through group relative policy optimization, which grades a model on task outcomes. Several methods can be chained together in a single multistage training run, and FAiE helps choose the strategy, set up rewards and diagnose problems along the way.

The training work feeds Oumi’s broader argument about model ownership. Most enterprise agents run on general-purpose models from frontier AI providers that aren’t optimized for a particular customer’s work. The company’s pitch since launching a commercial platform for custom models in March has been that a specialized model the customer owns outright can do better work at lower cost.

Group relative policy optimization is launching in private preview, and the Seattle company is taking requests for access. Oumi is a venture capital-backed startup that raised a $10 million seed round led by Venrock Management and Obvious Management Services in January 2025. The founding team includes former engineering leaders from Google LLC’s Gemini group.

Image: SiliconANGLE/GPT Image 2.5