Been thinking a lot lately about how most AI agents are essentially stateless — they execute tasks but have no real model of themselves or what they're doing. Built something to explore that problem. Colors is a local agent that maintains a persistent self-model and runs every action through a security classifier before executing. So before it touches your filesystem or makes a network call it actually reasons about whether it should. Still early but the behavior is genuinely different to what I've seen from other agent setups. Happy to share more if anyone's curious or has tackled similar problems. Repo is here if you want to poke around → https://github.com/thecolourfoundation/Color