Today's agentic software engineering stacks are built around frontier models. They are often the safest choice for achieving strong results. But as autonomous software factories scale across teams and parallel workflows, running every task on frontier models quickly becomes too expensive.
This talk digs into the hidden economics behind agentic software engineering: why benchmarks alone cannot tell us which model is appropriate for a specific task, where static model choices reach their limits, and how dynamic model routing could reduce unnecessary costs while maintaining the required quality.
We'll examine current approaches to using different models in agent factories, the trade-offs between static and dynamic selection, and open design questions such as when and at what level model choices should adapt.
