Hey HN, I wanted to show a quick hack I did extracting out some testing code I was writing to make AI agents easier to test.
My original goal was to make a shell where I could easily change some state, write a prompt, and get an AI agent (think "llm loop") to complete some task. I also wanted to support piping text files into said test harness.
What I realized while I was building this out was that I had inadvertently made an "agentic shell" that could run shell commands, so decided to extract out the code & open source it so that other people can use it and learn from it.
Currently it only supports Claude as the LLM, but there's work in progress to support OpenAI, Google, and maybe in the future local models.
0