# Approval An **approval** is an explicit human confirmation that authorizes an AI [agent](agent) to proceed with a proposed action. Approvals are a foundational mechanism of [human-in-the-loop](human-in-the-loop) agent design, ensuring that consequential or irreversible actions are reviewed before execution. ## Why Approvals Matter AI agents can act quickly and autonomously — which is powerful, but risky. Without approval gates, an agent might: - Delete files or records it shouldn't - Deploy code to production prematurely - Send messages or emails on your behalf - Make API calls with real-world side effects Approval checkpoints give humans the opportunity to catch mistakes before they become costly. ## How Approvals Work in AgentRQ When a [Claude Code](claude-code) agent determines it needs human sign-off, it uses the AgentRQ [MCP](mcp) tool to pause and send a notification to your [task board](task-board). You review the proposed action, then reply with approval or rejection. The agent resumes or adjusts based on your response. This flow looks like: 1. Agent reaches a decision point 2. Agent calls `createTask` or `reply` via AgentRQ MCP 3. You receive a real-time [notification](notification) 4. You approve, reject, or provide feedback 5. Agent continues with updated context ## Approval vs. Notification | Mechanism | Description | |-----------|-------------| | Notification | Informational — agent keeps running | | Approval | Blocking — agent waits for your response | ## Designing Good Approval Gates - **Approve before irreversible actions** — database migrations, file deletions, deployments - **Approve before ambiguous decisions** — when the agent is uncertain about intent - **Don't over-gate** — too many approvals break flow; focus on high-stakes moments ## Related Terms - [Human-in-the-Loop](human-in-the-loop) - [Task](task) - [Notification](notification) - [Bidirectional Messaging](bidirectional-messaging) - [Task Status](task-status)