Delegating Work to Agents: Goal, Verifier, Workspace

"How do you hand off a critical production migration to an AI agent and achieve a safer outcome than doing it by hand? In this talk, I will share a real-world case study: moving search architecture from a legacy Python service to a new path in under one month, deleting roughly 50,000 lines of old code, and reaching zero traffic on the old service without relying on users to discover regressions. Key Topics & Takeaways: - Defining a clear goal with a measurable pass condition instead of giving vague instructions. - Running continuous output verification against live traffic samples to catch ranking-order divergences and schema mismatches before code review. - Creating safe execution boundaries where agent-generated implementations can fail without production risk. - Where agent autonomy ends and human engineering authority remains essential."