AI Agent Management
Instructions
Define your agent's behavior, guardrails, model, and response style.
Overview
Open Build → Instructions to edit the directions your agent follows. General instructions describe its role and response behavior. Guardrails set boundaries for requests it should decline, clarify, or hand off.
General instructions
- Describe the customers and business the agent supports.
- Specify tone, language, and the amount of detail customers need.
- Explain how to use the knowledge sources and what to do when information is missing.
- Name the situations that require an action or human handoff.
- Avoid conflicting rules or promises the connected tools cannot fulfill.
Guardrails
Use the Guardrails editor for clear boundaries: information the agent should not disclose, claims it should not make, and situations that need a person. Test the boundary with both a request it should accept and a closely related request it should reject or escalate.
Model configuration
| Control | Purpose |
|---|---|
| Model | Choose from the models available in the workspace selector. |
| Temperature | Adjust the Reserved-to-Creative setting from 0 to 1 in steps of 0.1. Evaluate the effect on your own support questions. |
Compare before saving
- Edit General, Guardrails, or Model configuration.
- Choose Compare to evaluate the candidate settings alongside your saved settings.
- Use representative customer questions, including missing information and handoff cases.
- Review whether the change improves correctness and tool selection.
- Save the settings you want to use, then retest the deployed channel.
Channel instructions
A channel can use shared global instructions or its own configuration. When a channel is set to sync with global instructions, editing that shared text affects the other channels that sync with it. Review the channel's sync setting before saving.