research

Written in the open, published when they are true.

We study how agents should work inside a team, not beside one person. What we learn goes out as letters: short, specific, and with the mechanism named. The four below are being written now.

01AI made work single-player again.what a private chat window costs a team, and what a shared session gives back
02Permissions follow the room.an agent in a shared thread may only use what everyone there can see
03What one person teaches becomes a rule for everyone.corrections as skills, confirmed by a teammate, versioned and reversible
04The model is a supplier, not a landlord.why the tool you work in should not decide which model you use

Themes.

  • multiplayer AI in slack: what changes when the whole thread can see, correct and teach one agent
  • context shared between humans and agents: scope, versioning, attribution
  • the craft of designing, reviewing and shipping software with agents
  • models for collaboration: fine-tuned models built on kimi for candid, context-aware work without filler