Notes
🔍

AutoGPT

In March 2023 someone wired a language model into a loop, gave it a goal instead of a question, and let it run. The result, AutoGPT, was crude and oddly compelling — and it is the reason most people first heard the phrase "AI agent".

How it worked
You wrote down a role and a few objectives. From there the program ran unattended:

  • It broke the goal into a list of tasks and picked the next one.
  • It chose a command from a fixed menu — web search, read or write a file, execute Python — and ran it.
  • It fed the result back into the prompt, saving longer-term notes to files and a vector store, since the context window could not hold a whole run.
  • It repeated until it judged the goal met, or until you stopped it.

Why it mattered, and why it failed
The lessons were immediate and expensive. AutoGPT would pursue a plan long after it stopped making sense, re-run the same failing search, invent subtasks that drifted from the objective, and spend real money doing it. Nearly every discipline the field now treats as mandatory — step limits, token budgets, loop detection, structured tool schemas, human approval before irreversible actions — is a direct response to watching early AutoGPT runs. ReAct had already described the reasoning-and-action loop in a paper; AutoGPT showed a large audience what happens when you let one run free, and the frameworks that followed are largely about putting rails on it.

AutoGPT is the 2023 open-source project that popularised the fully autonomous agent: given a single goal, it prompts a language model to generate its own task list, executes steps with tools and file or web access, stores results in memory, and loops without per-step human input. It showed that the plan–act–observe cycle could run unattended, and equally exposed its failure modes — goal drift, repetitive loops, unbounded cost — which later frameworks were built to constrain.

Most AI tools wait for your next question. AutoGPT, released in 2023, was one of the first to be given a goal — "research this market and write me a report" — and then left alone to get on with it.

It wrote its own to-do list, searched the web, saved notes in files, and worked through the tasks one by one without being prompted again. For a few weeks it was everywhere, because nobody had seen software take that much initiative.

It was also unreliable: it got stuck repeating itself, wandered off the point, and ran up bills. Those failures taught the field what limits an independent assistant needs.