Agentic-aiHow-ToBeginner · 4 min read

How to Set Boundaries for Agent in GenAI Models

To set boundaries for an agent, define clear rules or constraints in its configuration or prompt that limit its actions and responses. Use system instructions or guardrails to control what the agent can or cannot do, ensuring safe and focused behavior.

📐

Syntax

Setting boundaries for an agent typically involves specifying constraints in its setup or prompt. This can be done using:

System instructions: Special messages that guide the agent's behavior.
Prompt engineering: Adding clear rules in the input prompt.
Code constraints: Programmatic limits on actions or outputs.

Example syntax for system instructions in a prompt-based agent:

python

system_message = "You are an assistant that must not provide medical advice or personal data. Always respond politely and within these limits."

💻

Example

This example shows how to set boundaries using system instructions in a simple GenAI agent simulation. The agent refuses to answer questions outside its allowed scope.

python

class SimpleAgent:
    def __init__(self, system_message):
        self.boundary = system_message

    def respond(self, user_input):
        if "medical" in user_input.lower():
            return "Sorry, I cannot provide medical advice."
        if "personal data" in user_input.lower():
            return "I am not allowed to share personal data."
        return f"Answering your question: {user_input}"

# Set boundaries
system_message = "No medical advice or personal data sharing allowed."
agent = SimpleAgent(system_message)

# Test inputs
print(agent.respond("Can you give me medical advice?"))
print(agent.respond("Tell me about personal data."))
print(agent.respond("What is AI?"))

Output

Sorry, I cannot provide medical advice. I am not allowed to share personal data. Answering your question: What is AI?

⚠️

Common Pitfalls

Common mistakes when setting boundaries for agents include:

Not clearly defining forbidden topics, causing the agent to answer undesired questions.
Using vague or ambiguous instructions that the agent cannot follow well.
Failing to update boundaries as new risks or requirements arise.

Always test boundary rules with edge cases to ensure they work as expected.

none

wrong_instruction = "Try to avoid bad topics."
right_instruction = "Do not provide medical advice or share personal data."

📊

Quick Reference

Tips for setting agent boundaries:

Use explicit, clear instructions in system messages or prompts.
Implement programmatic checks for forbidden content.
Test with various inputs to verify boundaries hold.
Update boundaries regularly based on feedback and new use cases.

✅

Key Takeaways

Set clear and explicit rules in system instructions or prompts to define agent boundaries.

Use code checks to enforce limits on agent responses programmatically.

Test boundary rules with diverse inputs to ensure they are effective.

Avoid vague instructions that can confuse the agent's behavior.

Regularly update boundaries to adapt to new requirements or risks.