Why I Fired My AI Agent

A computer science student opts out of assimilation.
Subscriber Only
Sign in or Subscribe Now for audio version

I am a college student, and students are using ChatGPT — a lot. We may say that we’re not, and for some that may be true, but it was a major player during finals week this spring. From helping to explain simple topics to translating foreign texts to writing essay outlines to writing full essays, ChatGPT has entered the classroom.

I’m a computer science major minoring in tech ethics. ChatGPT’s use, especially in my coding classes, is undeniable. It can debug faulty programs and write whole code projects. Even in my tech ethics class, students consulted ChatGPT to clarify ethical topics. There is a real possibility for plagiarism and cheating here. And so, initially, my main concern was how ChatGPT would change what it means to be a student. But now, after discovering the autonomous applications of ChatGPT, I am concerned with how they will change what it means to be a human.

With ChatGPT, which attracts most of the attention of AI news right now, it’s easy to conceive of it as simply a tool. You, a human, use it to ask questions and receive answers. But the shiniest new toys in AI are the applications of ChatGPT that act as “agents” such as AutoGPT and AgentGPT. Curious about their possibilities, I played around with each of them for a couple hours, and the results were incredible — scary but incredible. Whatever they are, they are more than simply a tool.

To simplify things a bit, imagine that someone were able to link together many chatbots, all asking each other questions, altering their next queries based on responses, and acting toward a common goal. Voilà! That is AutoGPT, which is described on its website as stringing together learned “thoughts” (with scare quotes) in order to “autonomously achieve whatever goal you set.” It “pushes the boundaries of what is possible with AI.”

For the average computer science student, the process of setting up AutoGPT is fairly simple. The open-sourced project is easy to clone from GitHub. With a couple of updates to already existing applications, four copy-and-paste commands, and one line to explain my choice of task, I was quickly running my new AI agent. AutoGPT is a free program, with one caveat: the paid API keys from OpenAI’s website that enable the chatbot’s GPT-4 capacities. After running AutoGPT for a couple hours to test it, my total rang up to less than 20 cents.

The possibilities seemed limitless. Currently looking forward to my travels to France for machine learning research, I decided to ask AutoGPT about planning restaurant reservations. From the simple line, “Find a high-rated restaurant available for dinner for two next Friday in Lyon, France,” AutoGPT created a to-do list of tasks and got started. Not only did the tasks include research of the best restaurants and their online availability, they also included using my email address to reserve the restaurant, with promises to check back in closer to the reservation date. Nowhere did I ask it to actually book the restaurant, but it did so anyway. It’s a good thing my new AI assistant knew what I really wanted to ask, I guess?

In accomplishing these tasks, I watched as the autonomous chatbot ran into dead ends with websites and altered the code until it ran successfully. For example, when queries of French restaurants on Google did not provide suitable results, the program updated the code it was using in order to try again with a new search engine and new queries, updating the file in real time and running it again.

This is the key difference from and radical innovation over ChatGPT: With ChatGPT, a user must continue inputting new queries when the search ends in a dead end, and must redirect the queries to achieve different outputs. With AutoGPT, I typed in one goal, then sat back and watched as the machine instructed itself, learned, reinstructed itself, and accomplished my goals. The chatbot — or, more accurately, autonomous agent — keeps going until it either completes the task or runs out of money on your API key from too many attempts. Or it will stop when instructed to — for now, at least.

Growing skeptical of the simplicity of using this autonomous agent, I dove deeper into pages of research on its website. The options for what can be done with AutoGPT seem to skew away from what the creators believe should be done. “Continuous mode” is available with just one command, but deemed “potentially dangerous” as it “may cause your AI to run forever or carry out actions you would not usually authorize.” What? The ability to carry out actions I would not usually authorize?

At this point, my fears heightened at what I unknowingly just deployed to destroy my computer. The website goes on to note that a virtual machine, a way to create an isolated testing environment on your computer, is strongly recommended to ensure “high security measures to prevent any potential harm to the main computer’s system and data.” Thankfully, my caution overpowered my curiosity, and by deleting everything I had downloaded and restarting my computer, I avoided any security breach or out-of-control AI agent. Phew.

AgentGPT, while similar in concept to AutoGPT, differs in its user interface. Aimed to “assemble, configure, and deploy autonomous AI agents in your browser,” it looks more like ChatGPT’s interface and so allows us to see the possibilities of widespread use of new personal AI agents.

As a website, rather than a coding project that requires downloading and multiple steps to set up, it comes with less safety risks and more free trials, but limited abilities for its tasks. When asked the same prompt about finding restaurants, AgentGPT simply did what I asked: it found a list of restaurants that were available this upcoming Friday, and produced with ease this information along with a suggested website to book “Michelin star level” reservations online.

Again, the only limiting factors in testing and using these programs were my 20 cents, my ability to read through instructions, and my patience in creating tasks and waiting for my new autonomous assistant to complete them, sometimes very slowly. As open-source projects, the range of tasks these and other AI agents can complete will continue to grow. Websites like There’s an AI For That, Baby AGI, and God Mode (where you can explore “hypothetical scenarios” and “think outside the box”) are all instances of ChatGPT embedded in applications to act as “agents.”

So far, they can gain access to your calendar and update it with events found in your email. They can make a new gaming app for you. They can create a new Twitter account and post updates according to certain goals. They can even use voice generation to finalize a pizza delivery order for you.

In using ChatGPT, you tell it what to do, over and over again, to complete your task. In exploring the embedded “autonomous” devices, I felt a shift of control. Where the experience of ChatGPT feels like a glorified web search, the experience of using the autonomous agents felt like I replaced a human interaction. As these devices grow in their task abilities, user interactions, and pervasiveness, they are changing what it means to be human. Are we content with letting these agents drive the human experience, or is now the time to step in and take the wheel?

More from the Summer 2023 symposium
“They’re Here… The AI Moment Has Arrived”

Clayton O’Dell, “Why I Fired My AI Agent,” The New Atlantis, Number 73, Summer 2023, pp. 43–45.
Header image: iStockPhoto / francescoch

Keeping humanity human

The New Atlantis is building a culture in which science and technology work for, not on, human beings.