Foundations for collaborative intelligence.

Sundial is a research lab in San Francisco studying human-agent collaboration. We build infrastructure for people to understand the work models do, steer them, and shape how they behave.

Model autonomy doubles every few months, but our capacity to review their work stays fixed. Closing that gap means making the work legible: understanding what was done, by whom, and why. Sundial builds instruments to keep that record and learn from it.

What we study

Q.I

Long-horizon verification

How can we efficiently verify agent work that would take humans weeks or months?

Q.II

Surfaces for understanding

What does the best surface look like for people and models to understand each other's work?

Q.III

Learning judgment

Can models learn calibrated judgment from records of past work and human feedback?

Q.IV

From context to weights

How do we tighten the self-improvement loop, so a model's learnings make it into its weights?

Instruments

Engraved horizontal sundial, light strokes on a dark plate

No. I · The editor

A minimalist editor for Markdown and LaTeX that lets you hand work to agents and stay in control. Every edit is attributed, and every change is yours to keep or reverse.

Meet the editor →Try the editor →Open the editor

No. II · Sun

A local-first system of record for human-agent work. It captures all human and agent actions in a folder, in enough detail to reconstruct how the work happened.

III

Join us

We are a small team hiring researchers and engineers. If you want human understanding to keep up with model autonomy, we should talk.

team@sundial.md