This repository is a curated prompt bank for collecting agent traces. It is not a trace dataset by itself. It is the input side: prompts that can be run through an agent harness, then logged into traces with tool calls, observations, and final answers. The goal of this release is practical:

keep prompts that work well in an agent harness remove prompts that assume hidden local state or user-private state expand browser and long-horizon tasks enough… See the full description on the dataset page: https://huggingface.co/datasets/kai-os/carnice-agent-trance-prompt-bank.