Anirban Chatterjee argues that software teams are moving from experimenting with AI coding tools to engineering repeatable systems, but faster code generationCode generation uses AI or another automated system to create source code from instructions, examples, schemas, or higher-level specifications. can create verification debt when quality, security and maintainability checks do not scale with output.
He cites studies that he says found a short-lived productivity increase alongside persistent static-analysis warnings, and that people often followed confident AI advice even when it was wrong. These research claims are reported from the talk and were not independently checked against the underlying papers.
The proposed response is zero-trust, multi-layered verification. Code should be checked consistently regardless of whether a human or AI wrote it, using independent computationalDeterministic software verification checks an output with a repeatable procedure that produces the same result for the same inputs and state. and reasoning-based methods with an auditable process rather than relying only on the same model or overloaded human reviewers.
The Guide, Verify, Solve loop gives agents relevant context and constraintsContext engineering designs how relevant instructions, evidence, state and tool information are assembled for a model's task. before coding, verifies work inside the agent loop and again in the wider CI/CD process, then lets agents remediate discovered issuesAn AI feedback loop occurs when AI outputs or their consequences become inputs that influence later model or agent behavior.. Anirban Chatterjee links this approach to bounded autonomy, earlier security findings, compliance evidence and consistent standards across teams and tools.
Watch on YouTube
