Increasingly, people can't explain what their agents did for them.
This would've sounded strange two years ago, but today it is ordinary. We are handing more of our thinking to AI agents than we ever have to any system, and we are doing it faster than anyone can verify.
For a long time the bulk of knowledge work was writing the work, and now it has become reviewing it. An agent can produce more in an afternoon than a person can read in a week, and the gap keeps widening, because the length of task an agent can finish on its own roughly doubles every few months while a human's capacity to review stays constant. Who is knowledge work for? For people...right?
Reviewing matters because agents make mistakes and hide them too. Review is also how we stay close to the work, and when we step back from it our own understanding fades.
For all of these reasons, we built Sundial. We want humans to stay at the center of the work. Sundial is a collaborative workspace for real work in docs, markdown and LaTeX, where humans and agents edit the same files live. It feels as familiar as the editor you already use, but with one difference: you can review any agent's or person's work at an exact, accurate level of granularity.
Here is how we are building it:
Version control for knowledge work: every change is tracked and tied to the agent turn that made it, reviewable at any level you choose, from a single edit to a chat session to a whole project.Any agent, and agent-native: Connect Claude Code or Codex in seconds and it joins as a collaborator.** You hold final sign-off**: the review panel shows exactly what was examined and at what level of detail, so you always know how much of the work a human has truly checked.
We have been working with select teams of researchers to submit papers to top conferences. In each case, multiple agents worked on Sundial to produce conference-quality research under human supervision.
Today, we are opening the private beta to more teams. If you would like to join us, sign up here and follow us on X.