Descriptions:
In this talk from the AI Engineer conference, Anirban Chatterjee of Sonar presents research and a practical framework for why AI-generated code requires systematic verification before it can be trusted in production. The talk opens with a Carnegie Mellon University study that used GitHub metadata to separate AI-assisted projects (primarily using Cursor) from traditionally written ones — finding a temporary three-month productivity spike followed by a persistent increase in SonarQube-detected static analysis warnings and code complexity that outlasted the productivity gains.
Chatterjee argues that this “verification debt” scales with application criticality. For experimental or internal tools with few users, the gap between AI output quality and acceptable quality is small enough to ignore. For production systems with adversarial users, large codebases, and compliance requirements, the gap becomes a material risk. The proposed solution is zero-trust, multi-layered verification — independent of how code was written — combining deterministic static analysis with LLM-driven review using a methodology deliberately different from the one that generated the code.
The framework Sonar introduces is ACDC (Agent-Centric Development Cycle), with three phases: Guide (providing guardrails, context, and constraints to the coding agent upfront), Verify (automated multi-layer code review covering quality, security, and compliance), and Solve (agentic iteration on identified issues). The talk is aimed at engineering teams moving from AI experimentation into AI-native production development workflows, and grounds its recommendations in observed data rather than vendor claims.
📺 Source: AI Engineer · Published August 09, 2026
🏷️ Format: Keynote Launch
