BabyAGI UI: A 2023 Walkthrough of the Task-Planning Interface

BabyAGI now has a web interface that makes its task loop easier to try. In this May 2023 walkthrough, I install BabyAGI UI locally and demonstrate it with an example objective. This is a historical tutorial: the [UI repository](https://github.com/miurla/babyagi-ui) was archived in October 2024, and current model availability, dependencies and setup may differ.

The original BabyAGI is an experimental task-management agent: give it an objective, and it creates, prioritizes and executes tasks in a loop. It is still basic and limited in what it can actually accomplish. Despite the name, this demonstration is not evidence of artificial general intelligence.

Yoshiki Miura’s BabyAGI UI offers a ChatGPT-like interface using Next.js, Tailwind CSS, Pinecone and LangChain. It can be deployed through Vercel or run locally. The UI is a separate project from Yohei Nakajima’s original BabyAGI; the latter has since evolved and its original version was moved to an archive.

For the local setup shown, I clone the repository with `git clone https://github.com/miurla/babyagi-ui`, enter it with `cd babyagi-ui`, and run `npm install` after installing Node.js. I copy the example environment file and start the development server with `npm run dev`. The repository now documents `cp .env.example .env`; consult the archived project’s instructions for its configuration rather than assuming the old demo remains current.

The README mentions environment variables and Pinecone. In this particular demonstration, I leave the environment settings unfilled and enter the OpenAI API key through the interface instead. That observation does not establish that Pinecone or other configuration is unnecessary for every version or feature; the repository’s later README specifies an index for its setup.

Once the server starts, I open the local address displayed in the terminal, here `http://localhost:3001`. The interface lets me choose from the models offered at the time and set a finite number of iterations or an unlimited loop. The model names, access and prices shown in 2023 should not be treated as current.

I choose GPT-3.5 Turbo, create an API key and save it in the interface settings. I say I will rotate the key shown during recording. API calls can incur charges, especially if the loop runs indefinitely, so monitor usage and stop the process when appropriate.

For the example objective, I enter “solve world hunger” with the initial task “develop a task list.” The model generates suggested tasks, such as researching economic, political and social causes, then prioritizes them. This demonstrates task generation, not a solution to world hunger.

The top task asks for an analysis of climate change’s effect on food production and distribution. The interface displays generated text, then creates and reprioritizes another list. A later task proposes a database of sustainable agriculture and food-security programs; the video does not establish that the system independently researched verified sources or built such a database.

I stop the run after seeing the cycle repeat. The task list and generated responses show the promise of a simple agent interface, but outputs need verification and real-world execution is limited. Continuous runs can create substantial API usage, so I recommend keeping an eye on the loop.