Short answer: Insert a gated approval step between the agent's "plan" phase and its "execute" phase. The agent submits the proposed action, a human reviews and approves or rejects it, and the agent only proceeds if approved. agentfabric.dev provides this as a native MCP server — zero infrastructure required.

Why add a human review step?

AI agents can take consequential actions: sending emails, modifying databases, publishing content, executing trades, calling external APIs. For high-stakes actions, you want a human to verify the agent's reasoning before the action is irreversible.

The right pattern: let the agent plan freely but execute only after human sign-off on the actions that matter.

Step-by-step implementation

1 Identify which actions need review
Not everything needs human review — that would defeat the purpose of automation. Define a risk tier: low-risk actions execute automatically, medium-risk get async review, high-risk are blocked until approved.

2 Connect agentfabric.dev
Add the MCP server to your agent config:

{
  "mcpServers": {
    "agentfabric": {
      "type": "streamable-http",
      "url": "https://mcp.agentfabric.dev",
      "headers": {"Authorization": "Bearer YOUR_TOKEN"}
    }
  }
}

Or via REST: https://rest.agentfabric.dev/v1/reviews

3 Agent submits proposed action for review

# Agent code (Python example)
import requests

def submit_for_review(title, content, context):
    resp = requests.post(
        "https://rest.agentfabric.dev/v1/reviews",
        headers={"Authorization": f"Bearer {TENANT_TOKEN}"},
        json={"title": title, "content": content, "context": context}
    )
    return resp.json()["review_id"]

review_id = submit_for_review(
    title="Send follow-up email to 500 users",
    content=draft_email,
    context=f"Triggered by task: {task_description}"
)

4 Human reviews in the reviewer inbox
The reviewer receives a notification (email, Slack, or the agentfabric.dev dashboard) with the full context. They can approve, reject, or modify the proposed action.

5 Agent polls and resumes (or stops)

import time

def wait_for_approval(review_id, timeout=3600):
    deadline = time.time() + timeout
    while time.time() < deadline:
        resp = requests.get(
            f"https://rest.agentfabric.dev/v1/reviews/{review_id}",
            headers={"Authorization": f"Bearer {TENANT_TOKEN}"}
        )
        status = resp.json()["status"]
        if status == "approved":
            return True
        if status == "rejected":
            return False
        time.sleep(10)
    return False  # Timed out — treat as rejected

if wait_for_approval(review_id):
    execute_action()
else:
    log_rejection_and_stop()

Using the MCP tool directly (for MCP agents)

If your agent already uses MCP (Cursor, Claude Code, Codex), the flow is even simpler — the agent calls the create_review_request MCP tool and waits for the get_review_result response:

How to Add Human Review to an AI Agent Pipeline

Why add a human review step?

Step-by-step implementation

Using the MCP tool directly (for MCP agents)

What agentfabric.dev provides

See also