Production traffic from frontier models is a golden data asset. If you can efficiently mine the traces, filter for quality, and fine-tune sm…

Production traffic from frontier models is a golden data asset. If you can efficiently mine the traces, filter for quality, and fine-tune smaller models on them, you get specialized performance at a fraction of the cost. This helps you build a continual learning flywheel: new traces, better datasets, better student models. One increasingly common flow teams are exploring here is Supervised Fine Tuning via Knowledge Distillation from their Traces Basically, we all use frontier models but they’re expensive. And really we just want them to be good at OUR tasks not EVERY task. These frontier model agents produce tons of good traces that show how to do a task So we can mine our Traces for these examples and use them to teach a smaller model how to do those tasks well The smaller model is way cheaper and faster to run and SFT is how we make sure it can match frontier models perf The important piece here is efficiently identifying which traces we can use to train a smaller model. Tools like online evaluators and running agents to read traces in bulk help build this data asset Our goal is to help every team efficiently mine their traces to improve their agents. SFT is one way to do this, will be posting more of how Production Data also helps you build RL Environments and replay your agent behavior to potentially push past frontier performance We want to help every team extract every last signal from every last agent Trace 🚀

Source: Harrison Chase (X) | 2026-05-29