Human-in-the-Loop Approval Workflows

The safest way to deploy high-leverage agents is by granting read-only autonomy, but requiring explicit user confirmation before executing write or delete operations.

Tiered Tool Permissions

Categorize every tool in your system into permission tiers:

Tier	Autonomy Level	Examples	Confirmation UI
Tier 1 (Read-Only)	Fully Autonomous	`search_docs`, `inspect_code`, `query_logs`	None (instant)
Tier 2 (Staged)	Autonomous Staging	`create_draft_pr`, `format_report`, `prepare_invoice`	Inline preview
Tier 3 (Destructive)	Explicit Human Approval	`merge_pr`, `send_email_blast`, `charge_card`	Modal confirmation

State Pausing and Resumption

When an agent requests a Tier 3 action, serialize the entire execution state into a database (PostgreSQL/Redis) with status WAITING_FOR_USER:

if (tool.isDestructive) {
  await db.sessions.update({
    where: { id: sessionId },
    data: {
      status: "AWAITING_APPROVAL",
      pendingAction: {
        tool: tool.name,
        args: tool.args
      }
    }
  });
  
  // Return interactive UI prompt to end-user
  return renderApprovalCard(tool.name, tool.args);
}

Once the user clicks Approve in the dashboard, the backend dequeues the execution, fires the tool, appends the result, and resumes the agent loop seamlessly.

Continue to Module 4: Evaluation Pipelines & Benchmarks.