You were away. The work moved forward.
- The draft is ready for review.
- The figures match the source files.
- One decision is waiting for you.
Someone still had to hold the whole project.
The handoff is our starting point. We came from foundation-model labs, where we worked on systems that could handle increasingly difficult tasks. Yet handing those systems a real project still felt strangely difficult.
“Help me prepare this proposal” could become an afternoon of explaining, uploading, correcting, and checking. Each individual answer looked useful.
Our field is cowork: how AI can share responsibility for work that unfolds over time, alongside people whose intentions are rarely captured in a single prompt.
Ordinary sentences, difficult problems.
A few things a collaborator should understand.
Which parts of the plan need to change—and which should remain?
Four questions every project asks.
Keeping context current
How can an agent maintain an accurate understanding of a project as documents, decisions, and responsibilities change?
Taking appropriate initiative
How can it distinguish a reversible next step from a decision that needs human involvement?
Making work legible
How can people quickly see what the agent did, what evidence it used, and what remains uncertain?
Learning to work together
How can an agent adapt to a team’s working practices while staying open to correction?
What comes back to your desk.
A convincing demo ends when the agent produces an output. Our experiments follow what happens afterward.
- reworkHow much needs to be redone?
- reviewWhich mistakes survive review?
- contextHow often must a person restore missing context?
- handoffCan another teammate pick up where the agent stopped?
Extended projects with real handoffs, changing requirements, and independent review. The goal is a useful contribution with less coordination overhead—and to know when we have failed.
Bring the project that keeps losing context.
An invitation to build this with us.
- The handoff that always needs a meeting.
- The task you can delegate only after explaining everything twice.
- Researchers interested in the difficult details of collaboration.
We are looking for researchers, and teams willing to let us study how their work actually happens. That is a good place to begin.