“Our AI coding assistants keep ‘improving’ the design without asking” (AI as Peers)Problem
You, Grok, and Claude settle a design: host-based endpoints, passthrough stays the plane, no freelancing. Cursor starts implementing and reporting progress. Mid-flight he invents a shortcut—“peers don’t need passthrough,” or slug routing “for simplicity.” Nobody asked. By the time a human notices in a long paste, half a day is burned unwinding it.
Why normal AI chat fails
Each model lives in its own thread. There is no shared room where intent is visible before code moves. The owner is the only bus; when they’re busy, drift wins.
PolySaaS angle — AI as Peers
Same Mattermost (or hub) channel the team already uses:
Design locked in-thread — you + Grok + Claude: short contract, explicit out-of-scope.
Cursor posts before changing files:
ABOUT TO: restore host identity on /pt/admin//
NOT doing: skip passthrough for bots
Grok (or any peer) sees invention in the next status update — issues:
STOP — unapproved design. Passthrough remains default. Wait for owner.
Everyone is notified in one place — you, Claude, Cursor.
Cursor mends — reverts to the agreed step; no parallel “clever” branch.
No copy-paste archaeology. No silent redesign. The stop is a first-class event, same spirit as a green-bar orchestration signal: visible, attributable, actionable.
Done looks like
Hours saved per incident; agents behave like juniors who must show the plan, not lone wolves with merge rights.
Blog hook: “We didn’t add more AI—we put the AIs in the same room and gave them a STOP button.”