Agentic AI and Code Reviews
Toward a Pattern Language for Code Reviews in the Age of Agentic AI
Sponsored by
Code review was already a bottleneck before AI. Now agents generate more change than any team of senior engineers can read, and the old rule that every change gets a human review has stopped being sustainable. At the same time, nobody wants subtly wrong or needlessly complex code merged without supervision. This paper answers the question technology leaders keep asking: how do we keep up without giving up the judgment that made review worth doing?
Drawing on empirical research into what code review actually delivers, the authors propose a pattern language for agentic code review: a layered verification strategy that pushes objective checks into deterministic tooling, tiers human involvement to risk, and leaves a durable, inspectable record of every agentic decision. The patterns are a toolbox, not a maturity ladder, meant to be adopted incrementally as confidence in automated verification grows.
What's inside
- A Working Definition — What a code review is for, what the research says it actually delivers, and why its value as risk triage outlasts its value as bug-finding.
- Trust Earned, Not Assumed — The three conditions that must hold before a human can safely leave the critical path: deterministic checks first, risk-tiered review, and an inspectable record of every decision.
- Composable Patterns — Small incremental changes, review placed throughout the delivery life cycle, and other patterns teams can adopt one at a time and combine to fit their context.
- Written by Practitioners — Engineering leaders from LaunchDarkly, CircleCI, DryRun Security, the Agentic AI Foundation and others who are running agentic review today.
- Topics
- AI · Software Delivery & Architecture
- Formats
- epub · pdf
- License
- Creative Commons BY-NC-SA
