Blog
Field notes from shipping AI-native systems
What we learned building agents that run in production — written by the engineers who were on call for them.
The 25% of agent interactions that consume 75% of your budget
Where token spend actually goes once an agent touches real systems — and the four controls that flatten the curve.
Restructure the work before you apply AI to it
AI compounds whatever process you point it at. Redesign first and it gets dramatically faster. Skip it and you run old bottlenecks at higher speed.
Write the eval harness before the agent
A golden set built from real cases is the only thing that turns a demo into something you can deploy on a Tuesday.
What we learned running a supervised agent for 14 months
Escalation rates, cost drift, model swaps, and the three incidents that changed how we grant capabilities.