Short answer
Auto-GPT is a free, open-source Python program released on 30 March 2023 that gives GPT-4 a goal and lets it decide its own next steps, such as searching the web, reading pages, saving notes and writing files, in a loop. It is an experiment, not a product: it asks for your approval before each action by default, it can be costly to run, and it often makes mistakes.
Follow-up question Should you try Auto-GPT yourself, and how do you stay safe?
Most people use GPT-4 one message at a time: you ask, it answers, you ask again. Auto-GPT, a new open-source project, removes you from the middle of that loop. You give it a name, a role and a few goals, and it decides for itself what to do next, carries out the step, looks at the result, and plans again.
The project was released on GitHub yesterday, 30 March 2023, by the developer Toran Bruce Richards, founder of the game studio Significant Gravitas. Its README describes it as "An Autonomous GPT-4 Experiment," and that last word matters. It is a demonstration of what might be possible, not a finished tool.
What happened
Auto-GPT grew out of an earlier version called Entrepreneur-GPT, and its original purpose is visible in the defaults. If you don't type your own instructions, the AI calls itself Entrepreneur-GPT and pursues goals such as increasing net worth and growing a Twitter account. The README includes a demo video dated 30 March showing it at work.
The code is free to download. To run it, you need Python (version 3.7 or later) and your own API key from OpenAI, the password-like code that lets a program use OpenAI's models and bills your account for what it uses. It runs in a terminal window, the text-only command line on your computer.
How it works: a loop of thoughts and commands
The clever part is a long set of instructions, called a system prompt, that Auto-GPT sends to GPT-4 at every step. It tells the model three things. First, its constraints: a memory limit of about 6,000 words and no help from the user. Second, a menu of commands it may use, including:
| Command | What it lets the AI do |
|---|---|
| Run a web search | |
| browse_website | Open a page and summarize it |
| memory_add / memory_del | Save or delete notes in long-term memory |
| start_agent / message_agent | Launch and talk to helper copies of GPT |
| write_to_file | Save text to a file on your computer |
| task_complete | Shut down when it decides the job is done |
Third, a strict reply format. Every answer must be JSON, a structured text format programs can read, containing one command to run plus the model's "thoughts": what it thinks, its reasoning, a short plan and a line of self-criticism.
The Python program then reads that reply, runs the chosen command, adds the result to the conversation, and asks GPT-4 for the next step. Here is how a simple run might go:
- You name the AI "TripPlanner," give it the role "an assistant that researches weekend trips," and enter a goal: "Find three affordable hiking spots within two hours of Boston and save a summary."
- GPT-4 replies with the command
googleand the plan "search, compare, write file." - The program runs the search and feeds back the results.
- GPT-4 chooses
browse_websiteon one result to read details. - After a few rounds, it chooses
write_to_file, thentask_complete.
By default, you stay in charge. After each proposed action, Auto-GPT shows you the command and waits; you type "NEXT COMMAND" to let it proceed. There is also a continuous mode that skips your approval entirely. The README is blunt about it: it is not recommended, may run forever, and may do things you would not normally allow.
Name: StudyBuddy
Role: an assistant that prepares a one-page study guide on any topic
Goal 1: Find three reliable beginner explanations of photosynthesis
Goal 2: Summarize them in plain language
Goal 3: Save the guide to a text file and stopWhy it matters
Auto-GPT is one of the first public examples of GPT-4 acting as an agent: a system that pursues a goal over many steps, uses tools, and decides its own actions instead of waiting for a person at every turn. Until now, most people have seen GPT-4 as a very capable conversation partner. This project shows a different shape: the model as the planning brain inside a loop, with simple programs as its hands.
It also exposes the weaknesses of that idea. Every step is a separate GPT-4 call, and OpenAI charges for GPT-4 by the token (roughly three-quarters of a word): about $0.03 per 1,000 tokens sent and $0.06 per 1,000 tokens received for the standard model. A long run with many steps adds up quickly. And because the model checks its own work, one wrong assumption can carry through many steps. The README itself warns that it could struggle with messy business problems in the real world and calls it "quite expensive to run."
Try it: Without installing anything, write a goal for an agent like the StudyBuddy example above. Then act as GPT-4 yourself: for each step, write down which single command you would choose and why. Count how many steps it takes, and note where a wrong guess early on would send the rest of the plan off course.
Should you try Auto-GPT yourself, and how do you stay safe?
If you are comfortable with a terminal and Python, it is an instructive experiment, and reading its prompt file alone teaches a lot about how to steer a language model. If not, there is no need to rush; the code is still changing quickly.
If you do try it, a few habits keep it safe and affordable. Set a monthly spending limit in your OpenAI account before you start, as the README itself advises. Leave continuous mode off, and read each proposed command before you approve it. Run it in a folder that holds nothing important, since it can write files. Never give it access to accounts or information you would not hand to a stranger.
Most of all, judge its results the way you would judge a keen but inexperienced intern. Check its sources, verify its facts, and expect that some runs will wander in circles. That is part of what this experiment is showing: how far a language model can get on its own, and where it still needs a person.
Sources
- Auto-GPT: An Autonomous GPT-4 Experiment (repository at commit 1f0905a, 30 March 2023) — Toran Bruce Richards (Torantulino / Significant Gravitas) on GitHub, March 30, 2023
- AutoGPT — Wikipedia, March 30, 2023
- GPT-4 — OpenAI, March 14, 2023