X Queue

1/ One Cursor agent for a whole web app is a mistake.

We run four. And before any code gets written, one of them does nothing but plan.

That single habit cut our rework more than any tool ever did.

Here's the exact setup 👇

2/ Start in planning mode, not code.

One agent maps the app, the data flow, and the edge cases first. No code yet.

Building blind is where rework comes from. Kill it early.

3/ Agent 1: The Planner.

Only job is architecture and sequencing. Writes the spec everyone else follows. No code.

4/ Agent 2: Backend.

Owns data models, APIs, logic. Reads the planner's spec instead of guessing.

Cleaner structure. Fewer surprises later.

5/ Agent 3: Frontend + landing.

Handles the interface only. Keep it away from backend so neither drifts into the other's mess.

6/ Agent 4: QA.

Reviews everything the others build. A dedicated critic catches what the builder can't see.

This is the agent most people skip. It's the one that saves you.

7/ Why it works: one agent juggling four jobs loses context and cuts corners. Narrow scope = sharper output, easier debugging.

But you still own the final call. We never ship a feature without a human reading QA's notes first.

Still running one agent for everything?