We study how agents earn their autonomy.
Our mission is to build truly autonomous agents: agents that can be trusted with work because they understand what their own actions cause.
Agents today execute well and predict well. Given a task and tools they finish it; given data they forecast. What they lack, in research and in industry alike, is any grasp of cause and effect in their own work: what changed because they acted, and what would have happened anyway. Without it, autonomy has to be granted by hand and policed by hand, and it never grows.
What we work on.
Cause and effect
How an agent can tell what its actions caused from what would have happened anyway, from the record of its own work.
Earned autonomy
How each piece of work can widen what an agent may do alone, and narrow it when the record says so.
Oversight
How people can review an agent’s work and correct it without slowing it down.
How we do it.
In deployment. The agents we build for finance keep the record this research needs, and the research decides what they may do next. Findings will be published here as they hold up.