This project is an attempt at creating an autonomous React.js Developer. It uses many of the concepts in BabyAGI, but extrapolates them slightly. Most notably by equipping the agent with tools to execute, and tailoring the prompts to the objective.
This project is just an experiment and not reliable what-so-ever. That said, hopefully it's a demonstration of some patterns that may be useful.
- Agent runs in the browser
- Defaults to using GPT-4, tested on GPT-3.5-turbo
- Uses Sandpack for the codebase/Renderer
- Uses Chakra UI for the UI
- Uses Langchain
- Based on "Task-Driven Autonomous Agent" model
Please add a OPENAI_API_KEY key to a new .env in the project root.
Now, run the development server:
npm run dev
# or
yarn dev
# or
pnpm devOpen http://localhost:3000 with your browser to see the result.
The pseudo-code for the agent is as follows:
- User enters an 'objective' (e.g. "Create a login UI")
- The Agent then creates an initial task list (the list is a JSON list of commands + inputs for the set of tools)
- The Agent then commences executing tasks in the task list, each tool interacts with the 'codebase' (we are using Sandpack)
- The Agent the determines any new tasks needed to be added to the task list, based on the output of the previous task
- The tasks are added to the task list
- The agent is then asked to prioritize the tasks
Tools are just functions that take a set of inputs, and act on the codebase.
Check out the /src/lib/agent/tools directory for more details.
Made at Planes Studio 🛫