One brain, every agent: extrapolating Cognition across hosts and teammates
Teach a skill in Claude Code and Cursor inherits it. Add a teammate and they start with the team's judgment already loaded. Here is how one brain scales across every host and person, keylessly, and why it is worth far more on a team than on one machine.
The value of a brain is not what one agent knows. It is what every agent inherits. A lesson trapped in one editor on one laptop is a note; the same lesson available to every agent and every teammate is infrastructure. Cognition is built around that second thing, so a skill learned anywhere shows up everywhere, with no re-keying and no copy-paste between tools.
Across hosts: the brain lives server-side
Claude Code, Codex, Cursor, and Gemini all connect to the same hosted brain. This works because the brain does not live inside any one editor; it lives server-side, and each host is just a client that calls it over MCP. Capture a skill while working in Claude Code and it is immediately retrievable from Cursor, because they are reading the same store.
"hosts": ["cursor", "codex", "claude", "gemini"]
The practical upshot is that you are not locked to one tool to keep your memory. Use the right editor for the task; the brain is constant underneath. A team that uses Cursor and Codex in different corners still shares one set of skills.
Keyless updates: why you never re-install
Hosted behavior updates for everyone at once, instantly, because it is server-side. There is no new key to issue and no re-install when the product improves. What can drift is the small local layer the installer wrote: the managed instruction block, your client_stub_version, and the cached tool list. Those self-heal:
"update Cognition" → cognition_update
This is the difference between "connected once" and "current forever." The local file is intentionally thin, just enough to load the latest hosted instructions each session, so the product can evolve without you reinstalling anything.
Across teammates: judgment as inheritance
When a skill is approved to the team brain, the next person's agent loads it before they start, with attribution intact, so they know whose judgment it is and when it was last confirmed. This quietly changes what onboarding means. Instead of a document a new hire skims and forgets, the team's hard-won judgment is something their agent already carries into the first task. The knowledge stops depending on who happens to be online.
Start with a brain, not a blank page
A new team does not have to begin from nothing. Install a prebuilt starter brain for your domain and customize as you learn:
"install a starter brain" → cognition_add_prebuilt_brain
Borrow a specific person's reasoning
Extrapolation is not only "everyone gets everything." You can apply one particular teammate's approach on demand, without them present:
"how would Alice approach this?" → query_employee_twin
This is useful exactly when you would otherwise interrupt someone: the person who always handles the gnarly migration is out, and you want their instinct, not a generic best practice. The brain has captured how they tend to decide, so you can apply it and keep moving.
Why it compounds on a team
Here is the part that makes Cognition worth more on a team than solo. Each host you add multiplies reach: the same skill now serves more agents. Each teammate who reports outcomes sharpens the ranking for everyone, because plasticity rewards distinct-agent reuse. Solo, a skill is a personal shortcut. On a team, it is shared judgment that gets stronger the more people lean on it. The network effect is real and it is the point.
next steps