An agent has to do more than produce a good answer. It must work with incomplete information, remember what matters, and recognize the limits of its authority. Useful agency should increase what people can do while preserving their control.
Questions we’re interested in
- How should an agent maintain useful state across long tasks?
- How do we make actions inspectable, interruptible and bounded by permission?
- How should agents recognize uncertainty, recover from failure and ask for help?
Our starting point
Our direction includes persistent memory, planning, tool use and learning from interaction. Everyday assistance and software engineering are potential applications of that work, with permissions and evaluation built into the research questions.
What this means today
We have not published an autonomous-agent release or agent evaluation results. Capability claims will need task definitions, failure analysis and clear deployment boundaries.
Safety & evaluation