At our company, agentic development is no longer just an experiment. For a few of our projects, we have shifted a large part of the development workflow to cloud-based AI agents that can work continuously.
In this talk, I’ll first give you a look at what that workflow actually looks like, including the tools we use, how agents fit into our development process, and how we make sure the code they produce still meets our quality standards.
But once you see the amount of AI involved, an obvious question comes up: how do we sustain the cost of running agents at this scale?
That is where the real topic of the talk begins.
I’ll share how we built an in-house system that optimizes our AI usage by deterministically routing different requests to the most appropriate models, instead of sending everything to the most powerful and expensive model available. I’ll also show how this approach has significantly reduced our AI usage and cost, backed by the analytics we collected along the way.
The goal of this talk is to share a practical approach to making large-scale agentic development economically sustainable, without simply compromising on the quality of the models you use.
// on_stage
