Descriptions:
Tom Howlett from Sonar examines what he calls the “enterprise quality gap”—the growing divergence between the speed at which LLMs generate code and the quality standards required to run that code in production enterprise environments. Presented at AI Dev 26 in San Francisco, the talk anchors its argument in a Carnegie Mellon University study of Cursor users that documented a 3-5x productivity spike in the first weeks of adoption, followed by rising static analysis warnings, increasing code complexity, and a productivity plateau by month three as teams struggled to debug code they hadn’t fully written.
Howlett draws a threshold at roughly 50,000 lines of code: below that, LLM-generated code with standard review processes works well, particularly for internal tooling and short-lived apps. Above that threshold—and especially for mission-critical systems that may run for years and face adversarial users—the current software development lifecycle doesn’t scale. He notes that he’s seen customers whose production codebases still include COBOL, illustrating just how long enterprise software must be maintained.
The practical portion of the talk outlines Sonar’s proposed verification framework: pairing test-driven development (TDD) and behavior-driven development (BDD) with static analysis to create a more complete, auditable, and deterministic quality gate for AI-generated code. Howlett cautions specifically against using one LLM to review another LLM’s code without additional verification layers, arguing this creates a false sense of confidence rather than genuine quality assurance. The talk is particularly relevant for engineering leaders at organizations scaling AI-assisted development beyond prototypes into regulated or high-availability production environments.
📺 Source: DeepLearningAI · Published May 21, 2026
🏷️ Format: Deep Dive
