Team Roles: Scout, Propose, Decide
Team Roles: Scout, Propose, Decide
When AI works with a team, someone needs to define who does what — or the AI will decide for you.
The Pattern
Most teams using AI fall into one of two traps:
- AI does everything — ships code nobody reviewed, makes architectural decisions nobody approved
- AI does nothing useful — restricted to autocomplete because nobody trusts it with real work
The fix is defining three roles and a simple flow.
Three Roles
Operator
The person who sees the big picture. They observe, prioritize, and direct where to look.
- Files issues, triages findings
- Manages branches and releases
- Sets direction — decides what matters, not how to implement it
Architect
The person who owns the code. Final authority on implementation.
- Reviews and merges all changes
- Decides patterns, conventions, and architecture
- When Operator and Architect disagree on code: Architect wins
Agent (AI)
The systematic worker. Analyzes, proposes, documents — but never ships alone.
- Scouts areas of the codebase for issues
- Writes PRs with tests and documentation
- Produces structured reports for humans to triage
The Flow: Scout, Propose, Decide
Operator observes → Agent scouts/analyzes → Agent proposes → Architect decides → Agent executes → Architect reviews
Escalation by level:
| Level | What | Who decides |
|---|---|---|
| L0 | Formatting, dead code, typos | Agent proceeds |
| L1 | Safe refactor, rename | Agent proceeds, Operator reviews PR |
| L2 | Behavior change, new utility | Architect approves before work starts |
| L3 | Architecture, new dependency | Architect approves before work starts |
The Scout Command
Tell your AI: scout <area> (e.g., scout security, scout performance, scout src/api/)
The AI produces a structured report:
- Area scanned — what files/modules were analyzed
- Findings — ranked by severity (critical → minor)
- Recommendations — proposed actions with escalation level
- Risks — what could go wrong if left unaddressed
The report is for humans to triage. The AI doesn’t act on its own findings.
Key Principles
- The architect’s patterns are the source of truth. The AI matches their style — it doesn’t “improve” it.
- The operator directs where to look, not what to write. They spot the forest; the architect handles the trees.
- The AI proposes, never ships. All changes require human approval before merge.
Give This to Your AI
Add this to your project’s agent instructions or CLAUDE.md:
## Team Roles
- Operator: observes, prioritizes, triages. Directs where to look, not what to write.
- Architect: final authority on code. Reviews and merges. Their patterns are the source of truth.
- Agent (you): scouts, analyzes, proposes. Never ship without human approval.
- Escalation: L0-L1 proceed; L2-L3 get architect approval first.
Created by Mats Ljunggren — LinkedIn