subagent-driven-development
Pass
Audited by Gen Agent Trust Hub on Oct 15, 2026
Risk Level: SAFE
Full Analysis
- [SAFE]: The skill adheres to security best practices for orchestrating multiple AI agents:
- Context Isolation: The instructions prevent subagents from inheriting the primary session's history or environment. They instead use a
SubagentContextPacketcontaining only the specific information required for the task, which limits the potential impact of indirect prompt injection and prevents context leakage. - Least Privilege and Role Separation: The skill enforces a strict separation of privileges where only the coordinator agent is authorized to mutate the Git state (staging, committing, branching). Implementer and reviewer subagents are restricted to editing files and providing read-only feedback.
- Mandatory Multi-Stage Verification: The workflow mandates two distinct review stages—specification compliance and code quality—before any work is accepted. This multi-layered verification process ensures that changes are both correct and safe before they are committed.
- Structured Escalation: The templates include explicit instructions for subagents to halt and request more information or more capable models when they encounter ambiguous requirements or architectural challenges, preventing unsafe execution based on assumptions.
Audit Metadata