agents

Jon Udell:

I dislike the phrase “human in the loop” because it cedes authority to the machines. Let’s flip the narrative. It’s our loop, we work the same way we always have, now we recruit agents to join the team. An agent-assisted process need not be a black box that takes in prompts and emits features.

Jon made me realize how easily I accepted the term “human in the loop” in the first place. Of course we’re in the loop, we made the damn loops! Just saying “human in the loop” implies there might one day be a loop without humans at all, which some people are trying to sell but not many people want.

I tend to agree with Jon that “agent in the loop” is a more appropriate disclosure, but if AI was ever capable of doing what big tech said it could, we might end up saying “agent(s) in the loop.” Not because we rightfully acknowledge and wish to disavow machine-centric language, but rather that we would want to acknowledge when an AI agent was keystone reviewer of a loop. As of right now, humans are still better than AI at a large variety of important review tasks, and until that changes “human in the loop” is the better name because human operators are the most valuable part.

The most profound insights about technology often come from direct experience rather than theoretical analysis. Last October when I gave my first-ever conference talk on agentic AI I emphasized process over code, specialized roles over general capability, and sequential collaboration over full autonomy. I was right about these architectural principles, but for entirely wrong reasons. The real limitations turned out to be more fundamental: accountability, security and an imperceptible line between capabilities and constraints.