Auto-GPT

Auto-GPT, now spelled AutoGPT, is an open-source program that let GPT-4 pursue a goal on its own by prompting itself in a loop: it planned a step, chose a command such as a web search or a file write, ran it, read the result, and planned again until it judged the goal done. Toran Bruce Richards, founder of the game studio Significant Gravitas, released it on March 30, 2023, two weeks after GPT-4 launched. It topped GitHub's trending list within days and turned the idea of an autonomous AI agent built on a large language model (LLM) from a research topic into something anyone could run. The AutoGPT repository has since passed 180,000 stars.

How Auto-GPT Works#

A user gave the original Auto-GPT a name, a role, and a few goals, such as researching a market and writing a report. Each cycle, the program built a prompt from those goals, a list of allowed commands, its constraints, and its recent history, and asked GPT-4 to answer in JSON with two parts. The thoughts part held the model's text, reasoning, plan, self-criticism, and a summary to speak to the user, a written-out form of chain-of-thought prompting. The command part named one action and its arguments: searching Google, browsing a web page, reading or writing files, executing Python code, or starting a helper GPT agent.

Before each command ran, Auto-GPT showed it to the user, who typed y to authorize it, y -N to authorize the next N commands, or n to exit. A continuous mode skipped these checks entirely, and the README warned that it could run forever or take actions the user would not normally authorize. For memory beyond the model's short context window, early versions stored past results as embeddings in a local JSON cache or in vector databases such as Pinecone, Redis, or Milvus, and retrieved the most relevant ones each cycle, while GPT-3.5 summarized long web pages.

What Auto-GPT Showed and Where It Struggled#

Auto-GPT made the ReAct pattern of interleaving reasoning with tool use visible to millions of people, and it inspired a wave of similar projects, including Yohei Nakajima's BabyAGI, which kept a running task list that the model created, prioritized, and worked through. Its runs also exposed the problems agent builders still design around. Agents repeated the same searches in loops, drifted away from the original goal, reported progress they had not made, a form of hallucination, and ran up API bills because every cycle called GPT-4 with a long prompt. The lack of a clear stopping rule and the risk of continuous mode showed why later agents run inside an agent harness with step budgets, approval rules, and an agent sandbox, and why teams measure them with agent evaluation instead of impressive demos. Models also improved at the core skill: OpenAI added native function calling in June 2023, and later reasoning models were trained with reinforcement learning to carry out long chains of tool calls.

Auto-GPT vs. Related Terms#

• Chatbots: Answer one message at a time and wait for the next. Auto-GPT generated its own next prompt and kept acting until it decided the goal was met.
• AutoML: Automated machine learning searches over models and hyperparameters to train the best model for a dataset. The names are similar, but Auto-GPT trains nothing and drives a pretrained LLM through general tasks.
• Robotic process automation (RPA): Replays scripted clicks and keystrokes that a developer recorded in advance. Auto-GPT chose its steps at runtime from the goal and each result.
• Agentic workflows: Arrange LLM calls and tools along paths a developer defines. Auto-GPT left the whole plan to the model, the fully autonomous end of the same spectrum.

AutoGPT Today#

The project has become the AutoGPT Platform, a hosted and self-hostable service for building agents as workflows of connected blocks on a visual canvas, with a marketplace of ready-made agents and a chat-based builder. The platform code is licensed under the Polyform Shield license, which allows personal and internal business use but forbids selling it as a competing hosted service, while the original standalone agent remains in the repository's classic/ folder under the MIT license, alongside the Forge toolkit for building agents and the agbenchmark harness for testing them. Goal-driven loops like Auto-GPT's now run mostly inside narrower products with fixed tools and permissions: in Ultralytics Platform, the Ask AI assistant turns a request into Platform actions, finding datasets, starting YOLO26 training runs, comparing their results, and proposing the next experiment, through the ul cloud command line authorized by the user's own API key.