AutoGPT's makers call its original autonomous loop a closed experiment

The conversation

AutoGPT was published on GitHub in March 2023. Its own original README, preserved in the classic/original_autogpt directory of the current repository, describes it as 'an experimental open-source application showcasing the capabilities of modern Large Language Models,' driven by GPT-4 to 'chain together LLM "thoughts" to autonomously achieve whatever goal you set,' calling it 'one of the first examples of GPT-4 running fully autonomously.' That looped design let the project set its own subgoals and call tools, such as file access or web requests, without a person approving each individual step.

What the documents show

The original README's own usage terms warn plainly that 'the use of the GPT-4 language model can be expensive due to its token usage,' and place responsibility for monitoring and limiting that cost on the person running it. The maintainers' current classic project README, retrieved 16 September 2026, now states that this original design 'is unsupported, and dependencies will not be updated,' describing it as 'an experiment that has concluded its initial research phase' and directing users toward a separate, actively developed AutoGPT Platform instead.

The system boundary

As originally published, AutoGPT's loop had no built-in step that paused for a person's approval before calling a tool; the goal-setting, prioritization and execution all ran inside the model-driven loop itself, with a human's role limited to setting the initial goal and, as the cost warning implies, watching the token bill. The project's own later classification of that design as a concluded research experiment, rather than as a maintained product, is itself the clearest boundary its maintainers now draw around what the original loop should be trusted to do.

Where it fails

Neither the original README nor the current classic-project notice claims the looped design reliably completed open-ended goals; the cost warning and later 'unsupported' notice are the maintainers' own acknowledgment that the approach needed either a person's active cost monitoring or, eventually, replacement. A builder looking at any unattended, loop-driven agent should read the absence of a reliability claim as informative, not as an oversight in the documentation.

• What stops a looped agent from repeating an expensive or harmful action before a person notices?
• Is there a hard budget or step limit enforced outside the model's own judgment about when to stop?
• If the maintainers themselves later called the design a concluded experiment, what does that imply for other unattended agent loops modeled on it?

AutoGPT's own repository is now the primary record of both what the original loop attempted and its maintainers' later judgment that the experiment had reached its limit.