michaelthwan / AutoDevGPT

Repository from Github https://github.commichaelthwan/AutoDevGPTRepository from Github https://github.commichaelthwan/AutoDevGPT

AutoDevGPT = Auto-GPT + development

Automated development using LLM technologies.

Incentive

  • I found that AutoGPT is kind of unstable to complete given goals. Therefore I limit the scope of the project to development tasks.
  • This project aimed to make a autoGPT like AI agent which can iteratively resolve development problems.

Roadmap

  • Development of a relatively complicated project can be roughly divided into 3 phases
  • Making a structure of flow, classes and methods without implementation - this part relies on human first.
  • Creating test cases (Test-driven development) for given methods - this part looks possible to be automated
  • Complete method implementation and run pass all defined test cases - this part is the first goal to be tackled.

AI agent setup

  1. Developer - 2 commands
  • create:
    • create a method from scratch. By experience it is possible for GPT to generate a method with lines of error
  • update:
    • update a method. If a method is close to completion, update it instead of re-creating from scratch
  1. Reviewer
  • criticise and suggest next action:
    • criticise whether an action is good. If not, roll back the code and re-do by the developer
    • Suggest a next action for developer.

About

License:MIT License