Creativity · Agent Protocol

AutoGPT (Original 2023)

The original AutoGPT, open-sourced by Toran Bruce Richards in March 2023, was the demo that made 'autonomous agent' a term civilians used. Give it a goal, walk away, and come back to find GPT-4 had been furiously making lists and Googling for hours. It was unreliable, expensive, and loopy — but it proved the category. Modern agent frameworks (LangGraph, AutoGen, CrewAI) are its intellectual descendants.

Protocol facts

Sponsor
Toran Bruce Richards / Significant Gravitas
Status
deprecated
Spec
https://github.com/Significant-Gravitas/AutoGPT
Interop with
Historical — original script replaced by modern AutoGPT platform

Frequently asked questions

Does the original AutoGPT still work?

The original 2023 script is preserved in Git history but no longer actively maintained in that form. Significant Gravitas has pivoted to 'AutoGPT Platform,' a different, more managed product. The original is a historical artifact.

Why was the original so unreliable?

It had no real planning layer, no memory consolidation, and no error recovery — just a loop that re-prompted GPT-4 after each step. With early GPT-4's coherence limits, it frequently got stuck in loops or chased irrelevant subgoals.

What did AutoGPT prove?

That a tight loop of LLM + tools could plausibly pursue long-horizon goals, and that there was real developer and public appetite for that capability. Every modern agent framework addresses one or more of AutoGPT's specific failure modes.

Sources

  1. AutoGPT GitHub — accessed 2026-04-20
  2. AutoGPT — original announcement — accessed 2026-04-20