What Production Knows: Closing the loop between agents and the systems they build
Coding agents write code fast, but they almost never hear back from production. Prompts go in, code comes out, and what the code actually does under real traffic stays out of the loop. Google's DORA research ranks software delivery instability as the second biggest effect of AI adoption. Developers feel faster while the system they ship into gets less stable.
Arun and May look at what it takes to ship agent-written code with confidence. Arun covers engineering controls that run as the agent works, on every file change, instead of gates that block it. May covers how runtime behavior from production can flow back to the agent and the reviewer: grading a PR against real traffic, catching a regression during rollout, and turning a regression into a fix PR an engineer can review.
Agenda
Controls that enable AI development at speed.
How AWS and Hud deliver production-grade AI.
Overview of Hud for Kiro.
Getting started.
Takeaways
Why "shift left" still assumes a linear pipeline, and what changes when controls fire on every file change.
How to grade a PR's risk from real traffic, so senior reviewers spend time on the risky changes.
How to catch a regression during a canary rollout by comparing against a learned baseline.
What changes when the engineer reviews a fix PR instead of hunting for the root cause.
Speakers: Arun Nallathambi, Builder, Applied AI Partnerships at AWS, and May Walter, Co-founder and CTO of Hud.
Register to get the recording.