AutoGPT starts the agent craze

Toran Bruce Richards, a game developer running a small studio called Significant Gravitas, released AutoGPT, an open-source program that wrapped GPT-4 in a loop: given a goal in plain English, it would break the goal into subtasks, decide what to do next, act — searching the web, writing and running code, reading and writing files — and feed the result of each action back into the next decision, continuing without a human approving each step. Existing uses of GPT-4 up to that point were essentially single-turn: a prompt, a reply. AutoGPT tried to make the model drive a multi-step task to completion on its own.

The project became one of the fastest-growing repositories in GitHub’s history, trending on the platform within days and passing 100,000 stars within weeks, alongside a wave of similar projects such as BabyAGI. It arrived a week after OpenAI opened ChatGPT to plugins and browsing, and together the two releases were widely read as the moment general-purpose language models started being pointed at open-ended tasks rather than single questions.

In practice, reviewers and users found AutoGPT unreliable. Because each step depended on the model’s own prior output, small errors compounded rather than self-corrected, and the program frequently fell into loops — repeating the same failed action, sometimes for hours, without making progress toward the stated goal. Each loop consumed paid API calls, so a stalled run could run up real cost for no result; commentators at the time put typical runs at several dollars for tasks that should have cost a fraction of that. AutoGPT’s limited context window also meant it could lose track of its own earlier reasoning on longer tasks.

AutoGPT did not become a widely used production tool, but it set expectations and a rough architecture — plan, act, observe, repeat — that later, better-resourced agent frameworks from OpenAI, Anthropic and others refined over the following two years.