AutoGPT was released on March 30, 2023. It is the project most non-specialists associate with the phrase “AI agent.” But “first” is one of the most contested claims in this field, and the record will not support a simple answer. Here is what the primary sources actually say.
What AutoGPT actually was
AutoGPT was an open-source Python project released by Toran Bruce Richards, founder of Significant Gravitas Ltd. It paired GPT-4 with a self-prompting loop, a web-browsing tool, file operations, and code execution. By April 3, 2023, it was the top-trending repository on GitHub. Within weeks, it crossed 100,000 stars — the fastest open-source project to reach that scale at the time.
That growth is what cemented “AI agent” in the mainstream vocabulary. But it does not, on its own, make AutoGPT first.
What shipped before it
Two things shipped before AutoGPT and have a stronger claim to “first” on different definitions:
BabyAGI (March 28, 2023). Yohei Nakajima posted a roughly 140-line Python script implementing a task-creation, task-execution, task-prioritization loop using GPT-4 and a vector database. It shipped two days before AutoGPT. It is the first widely shared autonomous LLM agent in the public record.
Adept ACT-1 (September 2022). Adept AI demonstrated a transformer trained to operate web browsers and software interfaces in response to natural-language instructions. ACT-1 was the first widely publicized transformer-as-action-agent — six months before either BabyAGI or AutoGPT.
And before any of them, the architectural pattern itself was already on arXiv.
The paper that defined the pattern
The reasoning-plus-action loop that AutoGPT, BabyAGI, and almost every subsequent agent framework uses is from a paper called ReAct: Synergizing Reasoning and Acting in Language Models, posted to arXiv on October 6, 2022 by Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. Read AutoGPT’s source and you can identify the ReAct pattern directly: a thought, an action, an observation, repeat.
So the engineering credit goes to Richards. The conceptual credit goes to Yao and colleagues.
And before that
The deeper lineage is older still. The idea that autonomous nodes can negotiate task allocation through a manager–contractor protocol is from Reid G. Smith’s Contract Net Protocol, published in IEEE Transactions on Computers in December 1980. Marvin Minsky’s Society of Mind (1986) gave the multi-agent view of intelligence its theoretical foundation. Michael Bratman’s BDI framework (1987) supplied the architecture. Stuart Russell and Peter Norvig’s textbook (1995) made the rational-agent abstraction the standard definition of AI itself.
None of those are AutoGPT, but each is part of why AutoGPT was a recognizable category and not a curiosity when it shipped.
So what’s the answer?
There isn’t one. There are several, depending on what you mean:
| Claim | Date | Project / Paper |
|---|---|---|
| First formal multi-agent coordination protocol | December 1980 | Contract Net Protocol (Reid G. Smith) |
| First widely demonstrated transformer-as-action-agent | September 2022 | Adept ACT-1 |
| First paper defining the modern LLM-agent loop | October 2022 | ReAct (Yao et al.) |
| First widely shared autonomous LLM agent script | March 28, 2023 | BabyAGI (Yohei Nakajima) |
| Project that put “AI agent” into the mainstream | March 30, 2023 | AutoGPT (Toran Bruce Richards) |
The most defensible position — and the one we take in the timeline — is that BabyAGI and AutoGPT shipped within two days of each other and should be recorded as effectively simultaneous origins of the modern open-source autonomous-agent movement. The conceptual origin is older, the engineering origin is one of two near-simultaneous projects, and the cultural origin is AutoGPT specifically.
“First” is doing a lot of work in any short answer. The honest version takes a paragraph, not a sentence.
Sources
- Yao, S. et al. (2022). “ReAct: Synergizing Reasoning and Acting in Language Models.” arXiv:2210.03629.
- Nakajima, Y. (March 28, 2023). “Task-driven Autonomous Agent Utilizing GPT-4, Pinecone, and LangChain for Diverse Applications.”
- Wikipedia: AutoGPT (release: March 30, 2023, by Toran Bruce Richards / Significant Gravitas Ltd.).
- Smith, R. G. (1980). “The Contract Net Protocol.” IEEE Transactions on Computers, Vol. C-29, No. 12.
- Adept AI (September 14, 2022). “ACT-1: Transformer for Actions.”
- See also the Agentic History timeline for the full chronology.