A video from VRSEN about the presentation of Devid - an AI software engineer. In it, the author demonstrates his open implementation of Devin, which has three main advantages: full access to the source code, training on real coding tasks, not just on GitHub Issues, and it is an agent system.
The author shows how Devid creates a website with the game "Game of Life", modifying HTML, CSS, and JavaScript files. Then he demonstrates how to import Devid and other agents into his own project using Docker containers. The author also describes how to set up planner agents, developer agents, and browser agents so that they effectively collaborate to perform tasks.
Finally, the author tests this agent system on a task of benchmarking several APIs, showing how agents can find documentation, execute code, and provide results.
So far, everything works quite mediocrely, although he blames OpenAI for this. If the documentation has many pages, errors occur. It is also not mentioned how many tokens these tasks consumed, it is simply stated that it is more efficient than in Devin.