BabyAGI

This is a museum piece and worth visiting as one. The 2023 script taught a generation of engineers what an agent loop looked like, and it was moved to a separate archive repository in September 2024, which is the correct end for a thing that did its job. There is nothing here you would run against work that matters.

Pick CrewAI if you want roles and a task queue that somebody maintains. Read the original if you want to understand where all of this started, then close the tab and use something current.

The failure mode is identity. The current repository holds a different framework from the one the name is famous for, built around storing and executing functions from a database. A reader who follows a three-year-old citation, or a package name in someone's requirements file, arrives at software that shares a title with what they were promised and nothing else.

Check what you actually installed before writing a line against it. What it does right: the original was moved to a separate archive rather than deleted, so the historical code is still readable and still attributable.

The architecture is three steps and worth stating precisely, because so much was built on it. 1. A task creation step proposes work from a result. 2. A prioritisation step reorders the queue. 3. An execution step runs the top task against a vector store used as memory. Nothing checks whether an executed task achieved anything.

That absence is the lesson. A loop that generates its own next task from its own last output, with no external signal, will happily run forever producing plausible work. Every serious framework since has added the step this one omitted.

The return on this project was never financial and it was substantial anyway. A weekend script became the vocabulary an entire category used to explain itself, and the person who published it acquired more distribution in developer circles than most funded startups manage in three years. That is a category-defining act with no cap table attached.

There is nothing to underwrite, no pricing to test and no acquirer to name. The projects that raised money on this idea are the ones worth analysing. Position: none, and a note that the most influential artefact in this category cost nobody anything.

My teams occasionally propose this because they recognise the name from an article, and the answer is the same every time. There is no vendor, no contract, no data-retention statement, no access control and no roadmap. The experimental framework currently under the name has no unattended mode either, so it could not be scheduled or monitored even if I wanted it.

Onboarding cost is irrelevant when the thing being onboarded has no owner. Engineers may read it on their own time. Not yet, and I do not expect that to change.

Nobody talks about what is actually in the repository now, which is a shame, because functionz is a genuinely odd idea: functions live in a database, dependencies between them are tracked, secrets are managed alongside them, and a dashboard shows the lot. It is a self-modifying program store with a web front end, and I can read every line of it.

MIT means I can take the parts I want. It is experimental, unfinished and going nowhere, and I still spent an evening in it happily. That is a fair trade for a project with no obligations left.