AutoGPT

AutoGPT builds tools for AI agents to carry out multistep work. Created by Toran Bruce Richards, founder of Significant Gravitas, it began as a standalone agent that generated its own prompts and selected subsequent actions toward a user’s goal. Its later AutoGPT Platform lets users assemble and manage workflows from models, external services, scripts, and conditional logic.[2][1][4]

Richards released AutoGPT on March 30, 2023. The original agent used an LLM to drive its actions and decisions: it generated intermediate “thoughts,” criticized its strategy, and chose what to do next without requiring a human prompt for every step. A May 25, 2023 Latent Space account reported that an automatic-code-generation video helped it go viral on April 1. By May, the article reported more than 132,000 GitHub stars, 37,500 Discord members, and one million downloads—distinct measures of project attention and distribution.[2][1]

At Agents Weekend, Richards presented a roadmap aimed at making AutoGPT accessible beyond users who could build and run a codebase. Plans included web and desktop interfaces, a mobile app, and an extensible plugin ecosystem called the Abilities Hub. The roadmap also proposed Challenge Driven Development: organizing improvements around goal-oriented tasks the agent struggled to complete. Concrete failures included getting stuck in loops, putting placeholders into commands, and issuing inappropriate commands. These were proposed improvements, rather than evidence that the failures had been resolved.[1]

Codium AI cofounder Itamar Friedman contributed testing work. In the May 2023 interview, he recalled using Codium AI to build tests for AutoGPT and finding problems through those tests. He described joining its management core team and speaking frequently with Richards. Friedman characterized AutoGPT as pursuing general-purpose agents, while Codium AI started with narrower code-integrity tasks. He also compared Challenge Driven Development with test-driven and dataset-driven development; the roadmap account connected the AI Maintainer project to one such challenge.[1]

A report by Matic on autogpt.net, published November 24, 2023 and updated November 8, 2024, said AutoGPT had secured $12 million in funding. The report does not establish investors, round type, valuation, or a closing date. It credited @Pwuts, @Swifty, @HunterBunter, and @ntindle as “Founding AI Engineers.”[6]

On September 24, 2024, Richards, identified in the announcement as CEO and creator, launched the AutoGPT Platform. The release offered GitHub self-hosting and a waitlist for a cloud-hosted beta. It defined an agent as a designed automated workflow, ranging from a linear sequence to a branching process with multiple decision points. Blocks represent actions: service connections, data processing, AI models, custom scripts, or conditional logic. Workflows could therefore combine AI tasks with ordinary software automation.[4]

The platform separated its server and frontend. The server contained core agent logic, infrastructure, and marketplace components; the frontend supplied agent building, workflow management, deployment controls, interaction, and monitoring. The launch described integrations with OpenAI, Anthropic, Groq, and Llama offerings, allowing model calls to become steps within a larger workflow.[4]

The launch also introduced Polyform Shield licensing for the new autogpt_platform folder while retaining MIT licensing for most of the repository. AutoGPT framed the decision as balancing community collaboration with a sustainable business model. Its announcement described the restriction as applying to competing products or services, including offering direct external access to the platform or wrapping it in a competing product.[4]

A June 30, 2026 announcement by product manager John Ababseh introduced @AutoGPT in Discord and Telegram. Teams could tag the bot with a request, let it plan and execute steps using connected tools, and receive results in a thread. The product supported shared-channel follow-ups, file exchange, background execution, and recurring schedules. The revised announcement described Slack support as awaiting app-review requirements.[5]

Ababseh reported that AutoGPT’s own team used recurring agents for a weekly product-metrics digest, a feature-flag audit, and release notifications.[5]

As of September 18, 2026, the repository described four platform interfaces: AutoPilot for conversational agent creation, Agents for supervising runs, costs, and actions needing attention, Marketplace for reusable agents, and Build for visual block editing. Agents could run on demand, on schedules, or from triggers. Hosted access was publicly available as a paid service with usage-based runs; self-hosting required users to supply infrastructure and model API keys. The original standalone agent remained in classic/ under MIT, alongside links to Forge for agent building and agbenchmark for benchmarking. The README claimed more than 185,000 GitHub stars, a measure of repository attention.[3]