Korva is a scroll of the internet’s best ideas.
The coworker model
The coworker model borrows how teams already work. The agent does a whole bunch of work, the person adds input, and the agent works again. They sync up at different cadences, look at each other's in-progress work, and trade feedback, which is a much more natural interface for working with agents.
One-on-one so far
Most agent work so far is one-on-one: a person and an agent, maybe with sub-agents, cut off from what colleagues are doing with theirs. At OpenAI there was a stretch when everyone sent screenshots of their threads with its coding agent, Codex, to each other on Slack to show how they had reached a number.
Multiplayer work
A screenshot is a poor way to collaborate, and it leaves open what the natural interface is for getting work done with agents together. The picture on the table is work that feels like a multiplayer game, with teammates steering their agents side by side while the agents handle more of the tactical rowing.
How did documents become multiplayer?
Real-time shared editing rests on operational transformation, a technique that C. A. Ellis and S. J. Gibbs described in 1989. It lets several people type into one document at once while the system reconciles edits that collide. Watching a colleague's cursor move through a shared draft then became ordinary office life. Google bought the startup Writely in 2006, and it became Google Docs.
Why did coding agents arrive first?
Code comes with built-in checks: a compiler either accepts it or does not, and tests either pass or fail. That gives an agent a clear signal about whether its work succeeded, which lets it loop and correct itself. Most office work has no equivalent pass-or-fail signal. Software also lives in plain text files, which language models read and write natively.
More from Lenny's Podcast: Product | Growth | Career
