© 2026 Run Loop LLC ·
Privacy
·
Terms
Loop
Home
User Case
Service
About Us
Docs
Contact
Notes
Projects
Archive
User Case
Last edited · Today
Decision recorded from this morning's design review: every agent in a Loop room boots into its own micro-VM cell. No shared state with peers. No shared state with the host. Why bother - When Foundry pushes a PR, we want to know that the build it ran is exactly the build the customer would ship. No env drift, no "works on my agent" excuses. - When Atlas opens an external page, that page can't see Ledger's queries. Period. - If something goes sideways, we kill the cell, not the room. How it actually works - Each cell is a snapshottable VM disk. Cold start under 2s. Warm resume around 80ms. - Suspend on idle. The cell goes to sleep with its FS, env vars, and in-flight tasks intact. - Branch a cell to A/B a decision: "what if Ledger ran this query against staging instead?" Two cells, same parent, you pick the one you like. What this unlocks for the user - "Fork this room" — try a different version of the work without disturbing the live one. - "Replay yesterday" — open the room exactly as it was at 4pm Tuesday, frozen in amber. - "Run 50 of these in parallel" — campaign mode, one cell per variant. TODO - [ ] Document the cell-to-cell IPC contract (currently in Jay's head). - [ ] Ship the "fork this room" gesture to the UI by next sprint. - [ ] Decide whether to expose snapshot history per agent, or just per room.