Request Approval

**Judgement:** Recommended for any environment where agents can modify real systems or data. It adds a strong human-in-the-loop safety gate around destructive or high-impact operations. **What it does:** - Uses Preloop’s `request_approval` MCP tool to pause before risky actions and ask a human to approve or deny. - Enforces a clear pattern: describe the operation, context, and reasoning, then wait for an explicit response before proceeding. **Key benefits:** - Substantially reduces the chance of accidental deletions, production misconfigurations, or unintended external side effects. - Comes with a concrete decision framework (ALWAYS / usually / don’t request approval) that makes agent behavior more predictable and auditable. - Simple interface (strings only) and broad notification support via Preloop (Slack, email, mobile) make it easy to integrate into existing workflows. **Risks and limitations:** - Depends on Preloop being correctly configured as an MCP server; misconfiguration or outages will block or degrade approval flows. - If overused for trivial actions, it can create “approval fatigue,” leading humans to rubber-stamp requests. - Only protects operations that the agent actually routes through this approval step; unsafe tools used directly can still bypass it if not integrated properly. **Recommended scenarios:** - Agents that can delete files/directories, touch databases, or manage cloud resources. - Any production or shared environment where stateful changes (deployments, config updates, CI/CD edits) are possible. - Workflows that involve sending emails, calling external APIs with side effects, or performing financial or security-sensitive actions. - Teams needing stronger governance and auditability around what agents are allowed to change and when.