When agents misbehave: How can we finally use AI at scale without breaking things

AI agents now have real access to real systems. They write and run code, call tools, touch production data and act on behalf of the people who deploy them. Most of what goes wrong isn't the model getting an answer wrong; it's an agent with more reach than anyone intended doing something nobody expected. In this opening keynote, Mat Velloso, Chief Product Officer at Docker, looks at what it actually takes to run agents at scale without breaking things — limiting what an agent can reach, keeping its actions visible and reversible, and building guardrails that let teams ship with confidence instead of locking agents down until they stop being useful.