AI Agent is here. How much should we let them do?

Should I set up a personal AI agent to help me with my daily tasks?

— Looking for help

As a general rule, I think relying on any form of automation in your daily life is dangerous when taken to extremes, and even potentially alienating when used in moderation, especially in personal interactions. A I have an agent That organizes my task list and collects online links for further reading? fabulous An AI agent that automatically texts my parents every week with quick life updates? terrible

The strongest argument for not adding more generative AI tools to your daily routine, however, remains Environmental impact These models continue during training and output generation. With all that in mind, I dug through Archives at WIREDPublished during the glorious dawn of this mess, we call upon the Internet to find more historical context for your question. After some searching, I came back convinced that you are probably using AI agents on a daily basis.

The concept of AI agents, or god-forbid “agentic AI,” is the current buzzword du jour for every tech leader trying to hype their latest investment. But an idea Auto Assistant Software dedicated to completing tasks is far from a new concept. Much talk of “software agents” in the 1990s mirrors the current conversation in Silicon Valley, where leaders of tech companies now promise an incoming flood of generative AI-powered agents trained to do online work on our behalf.

“One problem I see is people questioning who is responsible for an agent’s actions,” one read Wire interview With MIT Professor Patty Mays, originally published in 1995. “Agents in particular take too long at the machine or buy something you don’t want on your behalf. Agents will raise many interesting issues, but I’m sure we can’t live without them.”

I called Mace in early January to discuss how his view of AI agents has changed over the years. He’s as optimistic as ever about the possibilities of personal automation, but he’s convinced that “too naive” engineers aren’t spending enough time dealing with the complexities of human-computer interactions. In fact, he says, their recklessness could induce other AI winters.

“The way these systems are built, right now, they’re optimized from a technical standpoint, an engineering standpoint,” he says. “But, they’re not at all optimized for human-design issues.” He focuses on how AI agents are easily deceived or adopt biased assumptions despite improvements in the underlying models. And a false confidence leads users to trust answers generated by AI tools when they shouldn’t.

To better understand other potential pitfalls for personal AI agents, let’s break them down into two distinct categories: those that feed you and those that represent you.

Feeding agents are algorithms with data about your habits and tastes that sift through a wealth of information to find what’s relevant to you. Sounds familiar, right? any social media Suggestion engines filling a timeline with useful posts or endless ad trackers showing me those mushroom gummies for the thousandth time Instagram Can be considered as a personal AI agent. As another example from a ’90s interview, Mace mentions a newsgathering agent fine-tuned to bring back the articles he wanted. This sounds like my Google News landing page.