The beginning of the autonomous agent era
On 30 March 2023, a few weeks after the public release of GPT-4, developer Toran Bruce Richards (known by the pseudonym “Significant Gravitas”) published AutoGPT on GitHub. Within days, the project became the most discussed repository of the year and opened the season of autonomous agents: programs that, given a high-level goal, plan and execute the steps needed to reach it with minimal supervision.
The operational loop
AutoGPT is written entirely in Python and implements a four-phase cognitive loop: Plan → Think → Criticize → Act. In each cycle, the model plans the actions needed to progress towards the goal (Plan), reasons about the hypotheses at stake (Think), critically evaluates its own proposal (Criticize) and finally executes a concrete action (Act) — typically the use of a tool, writing a file or an HTTP request.
The loop continues until the goal is reached or explicitly interrupted. This architecture, while rudimentary compared to later structured agents, introduces the concepts of self-planning and self-criticism that will become standard patterns in the agent ecosystem.
Plugins and memory
AutoGPT introduced early two primitives that later became common: a plugin system to extend the agent’s capabilities with custom tools (browsing, code, file system) and an external vector memory (initially based on Pinecone, then extended to other stores) for context persistence beyond the model’s context window.
Licence and evolution
AutoGPT was originally released under the MIT licence. The project later split: the code core remains MIT, while the AutoGPT Platform — the commercial/managed version developed by the team — adopts a custom licence with specific restrictions. Beyond operational continuity, AutoGPT’s historical impact is clear: it demonstrated that a self-directed LLM, even with simple scaffolding, could carry out complex tasks — opening research and development around LangChain, BabyAGI and the entire 2023-2025 ecosystem.
Link: agpt.co