Auto-GPT — The Day 'AI Agent' First Went Viral

An AI agent is an AI system that, given a goal, decides its own steps, uses tools such as search or code execution, and chooses its next action from the results. What separates it from a chatbot that answers one question at a time is that the AI itself decides what to do and in what order, and keeps going until it is done. The idea broke out of research circles in the spring of 2023, and the trigger was Auto-GPT. Published on GitHub by Toran Bruce Richards, founder of the game company Significant Gravitas, the program took off with a demo video dated 30 March 2023 on its README and became one of the most watched repositories on GitHub.

Defining an AI agent

There is no single definition, but setting the most cited ones side by side shows the outline.

Anthropic puts it this way: "Agents, on the other hand, are systems where LLMs dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks." The dividing line, in other words, is whether a person writes the steps or the AI decides them. Auto-GPT showed the second kind in a form anyone could try on their own machine.

Auto-GPT arrives: 30 March 2023

The repository itself was created on 16 March 2023 under the name Entrepreneur-GPT, as an experiment in whether GPT-4 could run businesses to increase net worth. It began two days after GPT-4's release on 14 March 2023. By the end of March the README was titled "Auto-GPT: An Autonomous GPT-4 Experiment" and described the program as one of the first examples of GPT-4 running fully autonomously. Its demo video carried the date 30/03/2023.

The README of the time listed these features:

• Internet access for searches and information gathering
• Long-term and short-term memory management
• GPT-4 instances for text generation
• Access to popular websites and platforms
• File storage and summarisation with GPT-3.5

Once the user gave it goals, GPT-4 decided what to do, picked a command, ran it, read the result and decided again. By default it asked the user to authorise each step, but there was also a "continuous mode" that kept going without approval. The README warned: "It is potentially dangerous and may cause your AI to run forever or carry out actions you would not usually authorise."

BabyAGI: the other experiment that week

Two days earlier, on 28 March 2023, the investor and developer Yohei Nakajima had posted on his website a task-driven autonomous agent built from GPT-4, the Pinecone vector database and LangChain. GPT-4 played three roles: executing a task, creating new tasks from the result, and reprioritising the task list. The code went up on GitHub under the name BabyAGI on 3 April. In the same post Nakajima discussed the danger of a single-minded goal running away without ethical or resource constraints, citing the thought experiment of an AI that keeps making paperclips, and recommended a safety agent and ongoing monitoring.

Limits that showed at once

The limits were visible as quickly as the excitement. Auto-GPT's own README said it was "Not a polished application or product, just an experiment", that it might not perform well in complex, real-world business scenarios, and that it was quite expensive to run, so users should set and monitor their OpenAI API key limits.

Vice, describing both agents on 4 April 2023, reported research finding that language models tend to make things up further down a chain of subtasks, and relayed the AI researcher Andrej Karpathy's point that GPT models are unaware of their own limits, such as a finite context window and weak mental arithmetic, so that samples can go off the rails. Both Nakajima and Richards stressed that human supervision was still needed. IBM's explainer likewise lists drifting into irrelevant tasks, acting on hallucinated information and expensive API usage on large projects as weaknesses.

Afterwards: "agent" became the industry's word

Auto-GPT itself later turned into a platform for building and running agents, and as of October 2026 the repository had about 187,000 stars. The older parts are under the MIT licence, while the new platform is under the Polyform Shield licence, which forbids offering it as a competing hosted service.

What the 2023 Auto-GPT could not do, later products filled in one by one: Claude's computer use for operating a screen (2024), Devin, billed as an "AI software engineer" (2024), OpenAI Operator for running errands in a browser (2025), Claude Code for development in the terminal (2025), and MCP, the standard connecting agents to outside tools (2024).

Social impact and the sceptical view

Auto-GPT showed that, with an OpenAI API key and a little effort, people outside large companies could try an autonomous AI for themselves. At the same time the agent boom inflated expectations. In June 2025 the research firm Gartner forecast that more than 40% of agentic AI projects would be cancelled by the end of 2027 because of rising costs, unclear business value or weak risk controls. It described widespread "agent washing", the rebranding of existing products such as chatbots and RPA tools as agents, and estimated that only about 130 of the thousands of agentic AI vendors were genuine. In the same breath Gartner predicted that by 2028 at least 15% of day-to-day work decisions would be made autonomously. Expectation and disillusion have run side by side.

For the history of AI agents, the AI timeline and the open-source timeline are good starting points. For the language models Auto-GPT was built on, see the launch of ChatGPT and GPT-4.