The workflow I like with AI agents is that I give them the task and go about working on it. The only time I am actively chatting with one is when I am coding and bouncing ideas/implementations with Claude. The problem with just letting a coding agent run loose is that you may end up with some bullshit.

So to combat that, CodeBuddy implements a project wiki that stores all of the ideas, schemas, etc for a project. This is based on the LLM Wiki repo and serves as a knowledge base for the agent when they are working in that particular repo. After implementing this, I became more confident in the abilities of the coding agent to make reasonable decisions because there is some context as to what the preference is.

What has been everyone’s experience been with AI agents for non-coding work? I have built several of them for friends to automate pieces of their jobs (with and without company approval) and it looks to me that no one really wants to chat with these things while they work.

They treat it like a background process that tells them when the task is done. Sometimes they will talk to it to get the results via chat but that’s the limit.

What has been your experience with AI agents in the workplace?

For the life of me I couldn’t figure out how to get streaming http responses to work with yamux. It doesn’t matter anyways though. I can still tunnel regular http responses via yamux so the CodeBuddy project is still continuing.