The Agentic by Eigen event this week drew developers and researchers to hash out some of the thorniest problems facing artificial intelligence today. Chief among them: how to verify what an AI system actually does, who stays in control of autonomous agents, and why even the most advanced agents still fall short. The gathering made clear that the industry is hungry for systems that can be held accountable for their actions.
Verifiability takes center stage
One of the biggest themes was verifiability — the ability to check that an AI model's output matches its intended behavior. Without verifiability, users and regulators have no way to trust that an agent won't go off script. Participants at the event argued that current black-box models make verification nearly impossible, especially in high-stakes settings like finance or healthcare. The consensus: verifiability isn't a nice-to-have; it's a prerequisite for deploying agent-driven systems at scale.
Sovereignty and who's really in charge
Another worry that surfaced repeatedly was sovereignty. As AI agents become more autonomous, the question of who retains ultimate control becomes urgent. If an agent makes a decision that harms someone, who's responsible? The event highlighted a gap: most existing frameworks for AI governance don't account for the shifting boundaries between human oversight and machine autonomy. Sovereignty, in this context, means designing systems where humans can always pull the plug — and ensuring that design is enforceable, not just theoretical.
Agent limitations and the accountability gap
Even the most capable AI agents still bump into hard limits. They struggle with long-term planning, handling ambiguous instructions, and recovering from errors without human help. The Agentic by Eigen event showcased how these limitations aren't just engineering problems — they're accountability problems. Without reliable agents, no one can promise that a system will behave as expected. The event's undercurrent was a push for accountability baked into the architecture, not patched on after the fact.
What that looks like in practice remains an open question. The event didn't produce a blueprint or a set of standards. But it did sharpen the focus of a community that knows the status quo won't hold. The next few months will likely see more concrete proposals — and possibly disagreements — about how to make accountable AI agents a reality.
