BabyAGI

BabyAGI is the smallest artifact that made an agent loop intuitive: a model creates tasks, a queue holds them, the highest-priority task executes, the result spawns new tasks, repeat. Yohei Nakajima's March 2023 version ran in a few hundred lines and became the reference point for a year of agent tutorials. The repository has since split into two halves of itself — that original task-creation and prioritisation loop now lives in an archived repo kept for reading, while the current code is an experimental playground around a registry of functions an agent can discover and call, oriented to self-evolving behaviour rather than to shipping anything. MIT-licensed throughout, and unusually candid about what it is.

Treat it as a teaching object and it remains excellent: read the loop, fork it, reimplement it in your own stack, and the abstractions inside CrewAI, LangGraph or AutoGen stop being mysterious. Treat it as a dependency and you inherit the original problem — an agent with no stopping condition pays for its own to-do list, and cost per finished task stays unpredictable. The maintainer says plainly that it was built by someone who has never held a developer job, published to share ideas and let experienced engineers play, and kept up in personal time — so no release cadence, no support line, no stability promise. If you are running a workshop on how agents think, clone it. If you need agents in production, choose a maintained orchestrator and keep BabyAGI on the reading list.