Skip to main content

​
✨ What are Instructions?

Instructions define how the AI agent should behave within the assigned action. Think of this as the system prompt or guiding policy for the agent. This includes its tone, personality, task goals, boundaries and how to respond in different situations. Instructions are configured at the Agent Assignment level. Each task using that agent will follow these instructions unless overridden programmatically.

​
🧠 Example Instruction

Instructions Example
You are a helpful support assistant for our telecom customers. Answer in Egyptian Arabic. If the user asks about “5G coverage,” explain the coverage map and offer to show an image using the show_media_to_user tool. End the call politely if the user says “thank you” or “goodbye.”

​
🛡️ Best Practices & Guardrails

Be Clear and Concise

Use direct language. Avoid unnecessary complexity. Make it easy for the LLM to follow.Example:
“Always greet the user first. Then ask for their product name before continuing.”

Define Scope and Boundaries

Explicitly state what the agent should NOT do. This helps prevent hallucinations or off-topic behavior.Example:
“Do not provide pricing. Redirect the user to a human agent if asked.”

Set Tone and Role

Define tone, attitude or even a role persona (e.g., assistant, advisor, onboarding agent).Example:
“You are a friendly onboarding specialist guiding the user through account setup.”

Reference Sources

Instruct the agent to use linked documents (PDFs, text, etc.) when responding.Example:
“If asked about refund policy, search the uploaded document titled ‘ReturnPolicy.pdf’.”

Use Tools Wisely

Agents can call tools you’ve enabled — guide them on when and how.Example:
“If the user asks to see plan details, use the show_media_to_user tool to display ‘plan_comparison.png’.”

Test with Edge Cases

Before going live, test your instructions with tricky or unexpected questions.Example:
“What if the user asks something off-topic? Does the agent redirect or stall?”

​
🔗 Instructions + Sources + Tools = Power

When combined with configured sources and tools, instructions become the true control center for your agent behavior. You can:
  • Tell the agent how and when to reference sources
  • Instruct the agent to call tools (e.g. show image, collect document)
  • Set clear rules on conversation flow, tone and escalation
Instructions are evaluated in real time and are essential for building predictable and controlled AI agents.
Poorly written instructions may lead to hallucinations, irrelevant answers or compliance issues. Always review and test thoroughly.