I have started multiple projects using AI / Agent frameworks and have always been disappointed in the end. My current project I am implementing everything from scratch and I am much happier, I know where all the state exists and I do not have to spend hours trying to find how to extract some data from the agent loop which I need.
However today I was researching what I would deem to be "good" open source code in this area to try and find some interesting abstractions and noticed that nearly all the projects[0][1] are using Vercel's AI SDK for connecting to LLMs. Right now I have my own internal interface and am implementing a few providers (ollama, openai, anthropic).
So I wanted to see what the view of HN is, am I being stupid - is the AI SKD truly a good bit of abstraction and I should leverage it to save time?
- [0] https://github.com/sst/opencode - [1] https://github.com/VoltAgent/voltagent