Knowing when to stop an Agent Loop

Background

When using prompts in agent workflows like ReAct, it's important to know when to stop. Imagine you're making a web browser that can perform tasks, like ordering a 12 inch cheese pizza to your apartment. If your agent can order and pay for the pizza, that's great! But be careful: it might keep going and order more pizzas!! In these looping style workflows, it's really important to know when to stop.

Naive Approach

When I encountered this problem my first instinct was add a boolean property to the json response called complete. My prompt looked something like this:

Your goal is to "order me a 12 inch cheese pizza to my apartment"

Here are the tools avaiable:
	- CLICK - Click on a button
	- FORM_FILL - Fill out a form
	- BACK - Go back a page
	- GOTO - Go to a URL

You are on johnsofbleecker.com.

Response with the next action you want to perform. If you are done that just say complete: true. Here is the JSON format expected:

{
   tool: // One of the tools from above or empty is complete is true
   thought: // Explain why you are picking this tool - (I'll save this tip for another day)
   complete: // Boolean. Set to true if you are done
}

The goal of this prompt is to return JSON that looks like this when you need to use a tool:

{
   "tool": "CLICK"
   "thought": "Click on the order now button"
}

And then after the goal is complete something like this:

{
   "complete": true
   "thought": "Your pizza is orderd. The goal is complete"
}

While this sounds good on paper it leads to a big problem.

Problem

While this worked occasionally, it often leads to a major error. Even with GPT-4 I would receive JSON that looked like this:

{
   "tool": "Click on the order now button"
   "complete": true
   "thought": "Clcik on the order now button to complete the goal"
}

I tried different prompts or some conditional logic in code that checked for complete first but never solution worked well. It turns out there is a much easier way.

Solution

The solution is actually super simple - just turn complete into a tool. Instead of having a seperate boolean just create a new tool instead. Here’s what my prompt would look like:

Your goal is to "order me a 12 inch cheese pizza to my apartment"

Here are the tools avaiable:
	- CLICK - Click on a button
	- FORM_FILL - Fill out a form
	- BACK - Go back a page
	- GOTO - Go to a URL
	- CLOSE - CLose browser after the action is completed

You are on johnsofbleecker.com.

Response with the next action you want to perform. If you are done that just say complete: true. Here is the JSON format expected:

{
   tool: // One of the tools from above
   thought: // Explain why you are picking this tool - (I'll save this tip for another day)

}

This simple change lead to MUCH better performance. And if you want to fine tune a model this will be easier to fine tune since you are just focused on one property.