Who Decides if AI Code is Correct? Verification as the New Software Engineering

AI is writing code faster than engineering teams can verify whether it’s safe to deploy. The core bottleneck in software development has shifted: it is no longer "who will write the code?" but "who will decide if the code is correct?" While execution of checks can be offloaded to machines, defining what constitutes "correctness" cannot. This talk maps out the modern, stack-agnostic verification framework needed for agentic development—spanning specs, LLM evals, automated code reviews, CI guardrails, and production gates. We will examine where automation hits a wall, why reviewing clean-looking AI code introduces subtle hidden risks, and how to define agent autonomy based on action reversibility and blast radius. Featuring practical lessons from mobile deployment risks (where bad code bypasses checks to impact app store reviews) and findings from METR research on the perception vs. reality of developer velocity, attendees will learn how to structure verification layers and track metrics that prioritize engineering trust over PR volume.
Format: TECHNICAL TALK
Location: ROOM B
Speakers: Dmitri Nikitin
Level: Intermediate to Advanced
Audience: Registered participants