AutoGPT: 44,000 GitHub stars in a week, then the hard part of autonomous agents
AutoGPT, an open-source experiment in letting GPT-4 pursue goals on its own, became one of GitHub's fastest-rising projects in 2023. Early users hit loops, errors and high API costs. Its company later raised money and now sells a hosted 'team of AI experts'.
- Product
- AutoGPT
- Founder
- Toran Bruce Richards
- Launched
- 2023-03
- Model
- Open-source project plus a hosted subscription
Published · 2 min read · 6 sources
Key takeaways
- AutoGPT was released on 30 March 2023 by game developer Toran Bruce Richards and gained 44,000 GitHub stars in its first seven days, according to Jina AI.
- Early reviews found it prone to loops and errors, and Jina AI estimated about $14.40 in GPT-4 costs for a 50-step task.
- In October 2023 AutoGPT announced a $12M round from Redpoint Ventures; on 14 Nov 2026 the repo showed 188k stars and the hosted product listed Pro at $50/mo.
What it is
AutoGPT is an open-source app that uses OpenAI's models to act "autonomously": it asks and answers its own follow-up prompts until a task is done (TechCrunch). Users give an agent a name, a role, an objective and up to five ways to achieve it (TechCrunch). In 2026 its website sells a hosted "team of AI experts" (agpt.co).
Timeline
- 30 March 2023: Toran Bruce Richards, founder of video game company Significant Gravitas, releases AutoGPT, weeks after GPT-4 (Wikipedia).
- First seven days: the project gains 44,000 GitHub stars (Jina AI).
- April 2023: an AI researcher notes it has passed PyTorch in GitHub stars, 74k to 65k (TechCrunch).
- 14 October 2023: a post titled "AutoGPT Secures $12M Funding Round from Redpoint Ventures", linking to AutoGPT's X account, is shared on Hacker News (Hacker News).
- Checked 14 Nov 2026: the GitHub repository shows 188k stars and 45.9k forks (GitHub). The hosted product offers a 7-day free trial, "Then Pro at $50/mo" (agpt.co).
Why it worked
- Perfect timing. It was one of the first widely accessible apps to show off GPT-4's autonomous abilities (Wikipedia).
- A demo-able idea. Given a prompt like "help me grow my flower business", it could produce a plausible ad strategy and a basic website (TechCrunch).
- Open source on GitHub, which let developers try it and build on it; apps like AgentGPT and GodMode soon wrapped it in a browser interface (TechCrunch).
What to learn
- Stars measure attention, not usefulness. Jim Fan called it "a fun experiment" and warned that most "cool demos" were "heavily cherry-picked" (TechCrunch).
- Agents multiply model costs. Each step of a task calls GPT-4, and Jina AI estimated $14.4 for a typical 50-step task (Jina AI).
- Keep a human in the loop. Salesforce's Clara Shih: "it is especially important to include a human in the loop approach" (TechCrunch).
- A viral open-source project still needs a product. Three years on, AutoGPT's site leads with hosted "AI experts" and a paid plan (agpt.co).
Risks and what didn't work
- Loops. Jina AI listed "The Looping Quagmire: Why Auto-GPT Gets Stuck" among its main problems (Jina AI).
- Errors and forgetting. It can make errors, often doesn't recall how it completed a task, and struggles to break complex tasks into sub-tasks (TechCrunch).
- Setup friction. Early on it had to be installed in an environment like Docker and needed a paid OpenAI API key (TechCrunch).
- Funding detail. We could only open the Hacker News post about the $12M round, not the original announcement, so treat it as company-announced (Hacker News).
FAQ
Short answers drawn from the sections above.
What is AutoGPT?
AutoGPT is an open-source app that uses OpenAI's models to act "autonomously": it asks and answers its own follow-up prompts until a task is done (TechCrunch). Users give an agent a name, a role, an objective and up to five ways to achieve it (TechCrunch). In 2026 its website sells a hosted "team of AI experts" (agpt.co).
Why did AutoGPT work?
- Perfect timing. It was one of the first widely accessible apps to show off GPT-4's autonomous abilities (Wikipedia).
- A demo-able idea. Given a prompt like "help me grow my flower business", it could produce a plausible ad strategy and a basic website (TechCrunch).
- Open source on GitHub, which let developers try it and build on it; apps like AgentGPT and GodMode soon wrapped it in a browser interface (TechCrunch).
What can indie builders learn from AutoGPT?
- Stars measure attention, not usefulness. Jim Fan called it "a fun experiment" and warned that most "cool demos" were "heavily cherry-picked" (TechCrunch).
- Agents multiply model costs. Each step of a task calls GPT-4, and Jina AI estimated $14.4 for a typical 50-step task (Jina AI).
- Keep a human in the loop. Salesforce's Clara Shih: "it is especially important to include a human in the loop approach" (TechCrunch).
- A viral open-source project still needs a product. Three years on, AutoGPT's site leads with hosted "AI experts" and a paid plan (agpt.co).
What didn't work for AutoGPT, and what are the risks?
- Loops. Jina AI listed "The Looping Quagmire: Why Auto-GPT Gets Stuck" among its main problems (Jina AI).
- Errors and forgetting. It can make errors, often doesn't recall how it completed a task, and struggles to break complex tasks into sub-tasks (TechCrunch).
- Setup friction. Early on it had to be installed in an environment like Docker and needed a paid OpenAI API key (TechCrunch).
- Funding detail. We could only open the Hacker News post about the $12M round, not the original announcement, so treat it as company-announced (Hacker News).
Sources
Every link cited above, in order of first use.
- 1.TechCrunch (techcrunch.com)
- 2.agpt.co (agpt.co)
- 3.Wikipedia (en.wikipedia.org)
- 4.Jina AI (jina.ai)
- 5.Hacker News (news.ycombinator.com)
- 6.GitHub (github.com)
Related case studies
More in AI tools →Rabbit R1: 100,000 preorders for an AI gadget that reviewers said barely worked
Rabbit's $199 R1 sold on a CES keynote promising a 'Large Action Model' that would use apps for you. Reviews found most of it didn't work, and its interface turned out to run as an Android app.
Firecrawl: web data API for AI that raised $14.5M Series A while profitable
Caleb Peffer's Firecrawl turned Mendable's scraping pain into an API used by 350,000 developers. It raised a $14.5M Series A in August 2025 and says it was already profitable.
Supabase: a $10B backend, with most new databases started by AI tools
Supabase is an open-source backend built around Postgres. In June 2026 it raised a $500M Series F at a $10B pre-money valuation, and said more than 60% of new databases were being launched by AI tools.