Research at Red Rook AI

AI agents that work for days, learn new skills, and work together can take on much more useful work. Constant permission requests encourage rubber stamping, letting harmful actions slip through.

The risks grow when agents learn from their work. A mistake can become a lesson they repeat, and actions that seem harmless on their own can cause harm when combined. People need the power to instantly stop unsafe behavior and clearly audit what happened.

I’m building Keep to investigate how agents can take on more responsibility while people stay in control. The aim is to let agents work within clear limits, bring important decisions to a human, and keep understandable records so their work can be reviewed. Meaningful oversight requires AI whose actions people can examine and question.