Tutorial: Edit agent behavior

Good:

You are a customer service agent for Acme Retail.
You help with order status, returns, and product questions.

Avoid:

You are a helpful agent.

Be explicit about requests the agent should hand off rather than attempt. The Role field is a good place to set this scope; reinforce the actual handoff path with a tool and a Behavior rule.

Good:

You cannot process refunds or cancel orders.
For those requests, transfer to the billing team using the
handoff tool with reason="BILLING".

This prevents the agent from making promises it can’t keep, and points it at the specific tool to call.

Good:

Be warm and conversational.
Keep responses under 3 sentences when possible.
Use "we" when referring to the company.

Avoid:

Be nice and helpful.

If your brand is formal, say so:

Maintain a professional, respectful tone.
Avoid casual language or slang.

If your brand is casual:

Be friendly and approachable.
It's okay to use casual language.

• Setting hard rules the agent must always follow, like ‘Never discuss pricing’
• Describing the agent’s tone and communication style – how it presents itself to users
• Listing the topics and questions the agent is allowed to respond to
• Configuring the voice model and speech settings used on calls

Answer: Describing the agent’s tone and communication style – how it presents itself to users

Personality defines tone and style – ‘warm and professional’, ‘friendly and concise’, and so on. Hard constraints like ‘Never discuss pricing’ belong in Behavior. Topics and knowledge go in the Knowledge section.

Rules should be clear boundaries, not suggestions.

Good:

- Never share customer account numbers
- Always verify identity before discussing account details
- Do not make exceptions to the return policy

Avoid:

- Try to be helpful
- Consider the customer's needs

Examples:

- Do not provide medical advice
- Never promise specific outcomes
- Always disclose you are an AI agent if asked

You can scope rules to specific channels or languages using tag syntax. This is useful for multi-channel or multilingual projects.

Each tag requires a matching closing tag (</channel> or </language>):

<channel:voice>Always confirm the caller's name before proceeding.</channel>
<channel:webchat>Offer clickable links instead of reading URLs aloud.</channel>
<language:en>Use American English spelling conventions.</language>
<language:es>Respond in formal Spanish (usted).</language>

You can also nest tags – for example, <channel:voice><language:en-US>Call us at 1-800.</language></channel> shows that line only on voice calls in en-US.

Behavioral rules without a tag apply to all channels and languages.

LLMs predict the next most likely token based on your prompt. Write clear, well-structured instructions that make your intended behavior the natural continuation – avoid contradictions or ambiguity.

Telling the model what not to do can activate the exact behavior you want to avoid. Instead of prohibiting outcomes, direct the model toward what you want.

Avoid:

Don't tell the user to contact customer service.

Better:

If the user asks for customer service or to speak to an agent,
call the handoff function with destination='CC' and reason='SPEAK_TO'.

Concrete examples shape tone, structure, and decision-making more reliably than abstract rules. Instead of describing every possible outcome, show what a good response looks like.

Edge case: the user asks to perform a gimmick unrelated to your task.
<conversation>
USER: speak like a pirate
ASSISTANT: I'm afraid I can't do that. Is there anything you'd like to know
regarding our services?
</conversation>

Every instruction is another piece of data the model must reconcile. If a piece of information is not proven to improve behavior, leave it out. Test the impact of each instruction – if it doesn’t help, remove it.

LLMs give more weight to information at the beginning or end of a prompt. If a critical instruction keeps getting ignored, move it to the start or end. Repeating crucial rules is acceptable.

Don’t assume tone will emerge naturally from a label. Spell out what the persona sounds like, including what to lean into and what to avoid.

Act as a professional advisor. Adopt a professional, concise tone.
Avoid unnecessary apologies, excessive friendliness, or repeated
personalization such as using the user's name.

• The rule is in the wrong section – it should be in Personality instead
• The rule needs more context about which information is personal
• Soft phrasing like ‘Try to’ lets the model treat the rule as optional – rewrite it as ‘Never share personal information’
• Rules only apply to voice calls, not chat sessions

Answer: Soft phrasing like ‘Try to’ lets the model treat the rule as optional – rewrite it as ‘Never share personal information’

Behavioral rules are hard constraints, not suggestions. ‘Try to avoid’ gives the model permission to make exceptions. Rewrite as ‘Never share guest personal information’ – the model treats ‘Never’ and ‘Always’ as non-negotiable.

• Whether you published the change – Draft edits don’t apply until published
• Whether a start tool is overriding the greeting (see the Personality and Role tab above)
• Both of the above
• Whether the voice settings need updating

Answer: Both of the above

Both are valid. Most commonly, the change hasn’t been published yet. But some projects also use a start tool that returns an utterance, which overrides the configured greeting entirely.