Agnost AI Launches First Model Trained on Production Traces

Shubham, CEO of Agnost AI, posted that the company released its first model. He stated the model was trained directly on a customer's existing production traces. According to the post, this training produced better outcomes than frontier models for that customer's tasks. The announcement urged users to train their own models from agent logs instead of limiting those logs to debugging. Yohei Nakajima commented on the new site and shared the post. Robert Scoble also amplified the announcement from the company account. The startup describes its work as finding where agents fail and training improved models from those traces.