Research
What makes intelligent systems useful beyond a single interaction?
Our research spans the systems problems that emerge as AI operates across richer contexts and longer horizons: continuity, memory, adaptation, recovery, human control, verification, and coordination. We publish technical positions while they are still positions — measured claims wait for measurements.
Research agenda
07 areas
- Continuity
- How should intent and constraints survive time, failure, changing environments, and human intervention?
- Memory
- What should a system retain, revise, consolidate, or forget as its context develops?
- Adaptation
- How should a system revise its understanding when new evidence arrives, and tell outdated state from current state?
- Recovery
- How can a task resume after failure without repeating completed work or rediscovering known dead ends?
- Human control
- When should a system act, ask, wait, request approval, escalate, or return control?
- Verification
- How can behavior produced over hours or days be reviewed efficiently without hiding the decision that mattered?
- Coordination
- How should specialized agents exchange work while preserving ownership, intent, and system legibility?
How we publish
Positions
Arguments and design commitments. A position is not a measured result.
Engineering notes
Descriptions of systems we actually build, including architecture, implementation choices, failures, and revisions.
Evaluations
Measured results with methodology and enough context to understand what the result does — and does not — demonstrate.