Program · AGT·02
We build agents that plan, reason, and act, turning intent into reliable work across research, engineering, and everyday operations. The hard problem is not fluency but reliability over long horizons: staying coherent across hundreds of steps without drifting from the goal.
What we work on
A program is a small set of hard, coupled problems. These are the ones this group is working on now.
A model that is right ninety percent of the time is not ninety percent of a reliable assistant, because errors compound across a plan. We study the mechanisms of drift and design systems that notice when they are wrong, recover, and know when to ask rather than guess.
An assistant that acts in the world must expose its reasoning as structured, checkable steps rather than a black box. We build agents whose plans can be audited, replayed, and constrained, so trust is earned through transparency rather than assumed.
Real work happens through tools: code, search, data, and instruments. Our agents are trained to use them deliberately, verify the results, and compose them into workflows that hold up outside the demo.
Signals from the program
Selected writing from this program
Collaborate
We hire researchers and engineers who want to push one of these programs forward, and partners who want to put the results to work.
