Research
We study the systems layer: everything between a capable model and useful work that holds up over days. This is where we publish what we find.
Publications
The missing relation in agent state
Memory, checkpoints, and handoffs fail for the same reason. Systems keep their conclusions and lose what those conclusions rely on.
Read the note
Plans should carry intent, not payloads
If the size of a plan grows with the size of the artifact, the planner is doing the executor's job.
Read the note
The cost of checking
Delegation is bounded by verification.
Read the noteA capable model is one component of a capable system.
The model reasons inside a moment. Sustained work needs structure around it, and each part of that structure answers one question.
Seven open problems
Our research is organised around the parts of that structure that are still unsolved.
Continuity
Maintain intent and context across time.
The objective has to outlive any single model call. A restart, a failed dependency, or a correction from the person who assigned the work should change the plan, not erase it.
Memory
Retain, revise, and forget what should shape behavior.
Storing more context is not memory. The hard part is deciding what should influence the next step, and letting the rest fade.
Adaptation
Update understanding as context and evidence change.
A system has to tell outdated state from current state, and change course without discarding what still holds.
Recovery
Failure should cost progress, not the task.
Retrying is not recovery. After a failure, the system needs to know what was already attempted and where to pick up.
Human control
Define when systems act, ask, wait, or hand control back.
Ask about everything and it cannot carry real work. Resolve everything itself and it makes decisions it was never authorized to make. Useful autonomy needs explicit boundaries.
Verification
Make consequential AI behavior economical to review.
Watching every action stops scaling. We study review surfaces small enough to read and honest enough to trust. Read the note
Coordination
Let agents and people work together without losing intent.
More agents can mean more capability, or a system harder to understand than the task itself. We study how to add agents without losing legibility.
How we publish
Every note states which kind of claim it makes.
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.