Stop using multi-agent coding the wrong way.
Spawning 4 agents in parallel without coordination is not speed.
It's distributed chaos.
My Workflow
Start with a Scout (Agent 0)
Prompt: "I'll start 4 agents in parallel. Design task ownership for each agent and create cross-agent notes so they can communicate."
Then use Opus to design the task list:
- Define ownership boundaries upfront
- Create a shared cross-agent note format
- Scout handles conflict resolution and dependencies
- Scout doesn't write the code
- Scout designs the system
Without this, agents overwrite each other, duplicate work, break assumptions, and leave you merging spaghetti at the end.
Why Structure Matters
Manual multi-agent orchestration looks fast.
Structured orchestration ships faster.
When agents don't have clear ownership boundaries, you get:
- Conflicting changes to the same files
- Duplicated work across agents
- Broken assumptions that cascade into bugs
- A massive merge conflict you have to resolve manually
When you invest 5 minutes in Scout-designed task ownership, each agent knows exactly what to touch and what to leave alone.
The Cross-Agent Note Format
Each agent should produce structured notes that other agents can consume:
## Agent 2 — API Layer **Owner:** All files in /api/* **Dependencies:** None upstream **Outputs:** REST endpoints at /api/v1/* **Do not touch:** Frontend components, database schema **Status:** Completed auth endpoints, working on data endpoints
The Scout reads these notes before resolving any conflicts. It's the coordinator, not the coder.
Result
Sessions that used to end in merge chaos now ship in one clean pull request.
How do you run your multi-agent sessions?