How to Intermediate · 3 min read

How to add human approval to AI agent actions

Quick answer

Add human approval to AI agent actions by intercepting the agent's proposed actions and prompting a human reviewer to confirm or reject them before execution. Use a control loop in your code that pauses for human input after the AI suggests an action, ensuring oversight and safety.

PREREQUISITES

Python 3.8+
OpenAI API key (free tier works)
pip install openai>=1.0

Setup

Install the OpenAI Python SDK and set your API key as an environment variable for secure access.

bash

pip install openai

Step by step

This example demonstrates a simple AI agent that proposes actions, then waits for human approval before proceeding.

python

import os
from openai import OpenAI

client = OpenAI(api_key=os.environ["OPENAI_API_KEY"])

# Function to get AI agent's proposed action

def get_agent_action(prompt):
    response = client.chat.completions.create(
        model="gpt-4o",
        messages=[{"role": "user", "content": prompt}]
    )
    return response.choices[0].message.content.strip()

# Main loop with human approval

def run_agent_with_human_approval():
    prompt = "Suggest the next action for the agent in plain text."
    while True:
        action = get_agent_action(prompt)
        print(f"AI proposes action: {action}")
        approval = input("Approve this action? (yes/no/exit): ").strip().lower()
        if approval == "yes":
            print(f"Action '{action}' executed.")
            # Here you would add code to actually perform the action
            break
        elif approval == "no":
            print("Action rejected. Asking AI for a new action...")
            prompt = "Suggest a different action for the agent."
        elif approval == "exit":
            print("Exiting agent loop.")
            break
        else:
            print("Invalid input. Please type 'yes', 'no', or 'exit'.")

if __name__ == "__main__":
    run_agent_with_human_approval()

output

AI proposes action: Send an email to the client.
Approve this action? (yes/no/exit): no
Action rejected. Asking AI for a new action...
AI proposes action: Schedule a meeting with the client.
Approve this action? (yes/no/exit): yes
Action 'Schedule a meeting with the client.' executed.

Common variations

You can implement asynchronous approval workflows using async Python and webhooks, or integrate approval UIs in web apps. Different models like claude-3-5-sonnet-20241022 can be used for improved reasoning. Streaming responses allow partial action previews before approval.

Troubleshooting

If the AI repeatedly suggests undesired actions, refine the prompt or add constraints.
If input hangs, ensure your terminal supports input() or switch to a GUI/web interface.
For API errors, verify your API key and network connectivity.

✅

Key Takeaways

Intercept AI agent actions and require explicit human approval before execution.
Use a simple input loop to confirm, reject, or exit agent actions interactively.
Customize prompts and models to improve action relevance and safety.
Consider async or UI-based approval for scalable human-in-the-loop workflows.

Verified 2026-04 · gpt-4o, claude-3-5-sonnet-20241022

Verify ↗