Agent Nation: How Autonomous AI Is Rewriting the Rules of Society—and What We Can Do About It

I've spent close to two decades studying how people look for information and how systems decide what to show them. Search, recommendation, ranking. For most of that time, the ethical questions in my field had a particular shape: what gets surfaced, what gets buried, whose perspective ends up on page one. Serious questions. But the system was still handing you options. You chose.

Somewhere around 2022 and 2023, that stopped describing what I was seeing. The systems weren't presenting options anymore. They were closing loops. Not "here are candidates you might consider," but "these are rejected, these move forward." Not "here's information about your claim," but a determination. The interesting part is that this wasn't announced. There was no launch event for the end of the human in the loop. It happened one procurement decision at a time, because someone had a backlog and the software was cheaper than three more staff.

That's the shift the book sits on top of, and it makes a lot of the existing literature slightly off-target. Cathy O'Neil's work on algorithmic bias, which I admire, diagnoses flawed models producing unfair outcomes. Still true. But a biased model is a static object you can audit. An agent adapts, acts in sequence, and produces consequences that nobody specified in advance. Mustafa Suleyman writes about containing powerful future technology. Verity Harding argues for public participation in setting AI's direction. Both matter. Both are aimed slightly downstream of where I think the pressure is. The agents already deployed in ordinary bureaucratic settings are doing more to reshape people's lives right now than anything in a frontier lab.

What pushed me from "this is a research interest" to "this is a book" was watching who gets caught in the gap. I've spent time with judges, federal agencies, and people building these systems inside companies. What struck me was that almost nobody in those rooms was acting in bad faith. The engineers thought the policy people were handling oversight. The policy people assumed the technical controls covered it. The vendors pointed at their clients. The clients pointed at the vendor.

And in the middle of that polite circle of deferral is a person who got denied something and can't find out why.

I don't think you fix that with better models. Better models are coming regardless. You fix it by building institutions that can answer a simple question on demand: who decided this, and how do I contest it. We have a long history of doing exactly that for other powerful systems. We just haven't done it here yet.

Ongoing thread. More from Chirag Shah to follow.