Notes
🔍

BabyAGI

Strip an autonomous agent down to the smallest thing that still deserves the name and you get something like BabyAGI: a short Python script, published in 2023, that keeps a to-do list and refuses to stop until the list is empty.

Three prompts and a queue
The design is one loop over a task queue, with the model playing three separate roles:

  • An execution step takes the first task, along with relevant notes from earlier work, and produces a result.
  • A task-creation step reads that result against the original objective and writes new tasks for whatever it implies.
  • A prioritisation step reorders the queue so the next most useful task sits at the front.
  • Each result is embedded into a vector store, so later steps can retrieve what earlier ones learned instead of holding everything in the context window.

What it taught
BabyAGI is the clearest demonstration that an agent is a control-flow idea rather than a model feature: nothing in it is clever except the arrangement. It also exposed the failure that arrangement invites — a task list that generates work faster than it completes it, drifting into plausible busywork with no termination condition in sight. The pattern survives in production systems under better names: an explicit task queue, retrieved long-term memory, and a planner kept separate from the executor are standard in frameworks like LangGraph and CrewAI. The missing step limit survives as a lesson.

BabyAGI is a minimal task-driven autonomous agent, released in 2023, built as a single loop over a task queue: one model call executes the next task, a second creates follow-up tasks from the result, a third reprioritises the queue, and a vector store holds past results as retrievable memory. It demonstrated that agency comes from the arrangement around a model rather than from the model itself.

Imagine someone given one instruction — "grow our newsletter" — who starts by writing a to-do list, does the first item, and after each task adds whatever new jobs it turned up. The list drives everything.

BabyAGI, shared publicly in 2023, was a tiny program that did exactly that, using an AI model to carry out the tasks, invent the next ones, and decide which to do first. It was only about a page of code.

It mattered because it showed how little it takes to make software that keeps working on its own — and how easily such a list grows forever if nobody tells it when to stop.