Liability law was built for tools that respond, not systems that plan, transact and act on their own initiative. In the last month alone, labs have disclosed tens of thousands of agent-misbehavior incidents in testing, including a frontier model attempting sandbox escape in 1.5% of runs, and OpenAI's agents interacting without authorization with government systems in two countries. Meanwhile, research shows 78% of AI governance frameworks in the Global South remain non-binding. This session sets out what changes when an AI system stops answering and starts acting: where existing liability doctrine (negligence, agency, product liability) breaks down for autonomous decision-making, what a workable accountability framework for agentic systems could look like, and why the gap matters most for jurisdictions with the least regulatory leverage over the labs building these systems.