Data Tech 2026

As corporations shift from standalone AI models to multi-agent systems—where a "supervisor" agent hands off tasks to specialized teammates—AI governance is evolving from worrying about individual model flaws to tackling the trickier risks of coordination. Think of it like a corporate hierarchy: delegation creates blind spots (agents handle private data, hidden reasoning steps, and tool results) and can spark misaligned motives, leading to "agency loss"—the frustrating gap between what you intend and what the system actually delivers. Drawing from recent research, I'll frame these dynamics as classic principal-agent problems from economics and game theory, where tools like addressing moral hazard, countering adverse selection, and crafting smart mechanisms offer sharper insights than rigid checklists ever could.

In this talk, we'll unpack how everyday MAS governance tactics—everything from evaluation suites and risk grading to release checkpoints, live monitoring, and escalation protocols—map onto game-theoretic levers that reshape the incentives and equilibria. You'll walk away with a hands-on framework for treating multi-agent setups as strategic games: designing payoffs, verifications, and oversight to curb deception, boost dependability, and turn agentic AI into something truly decision-ready.