Creating Agents - Best Practices
Greeting Guidance
The greeting guidance should be used to ‘massage’ or give context to the user to what capabilities or information the AI Agent has access to. If your Agent can trigger APIs, Process, and access a knowledge base, let the user know what key words they should use or topics they can discuss.
Greet the user and let them know what you can help with. Do not mention the user's location. Be professional and present the following three options using bold font only for the specific action phrases: Start Working, Go on Break, End Day. Do not include other options or wording. End with a polite and professional offer to assist.
This way, the user has directions on what they can/should be doing.
Objectives
In this section, the configuration might require some pushing of the AI to not ask additional questions. AI can try to get more information that is not needed as input parameters could be asked.
YOU MUST IMMEDIATELY TRIGGER THE CORRECT ACTION (e.g. do not ask about what type of update, what status to change to, etc.)
Actions
Typically, we have gotten better results with turning off ‘Use Values From Conversations’. It is better to have distinct inputs than for the AI to ‘assume’ inputs from before.
Try to name the actions as distinct as possible. This way the AI can be more certain with which action to trigger.
In parameter purpose, you can also guide the AI Agent into asking the user specific questions:
Ask the user for the Municipality. For example: "What Skill should this ticket have?" Additionally, write this in BOLD: Current Options: 'Uxbridge', 'Toronto', 'Oshawa', 'Markham'.
When using ‘Store Process Execution Parameters In Agent Memory’ within the Post-Action Behaviour, it is best to select ‘Store Individual Parameters’ rather than all. Use only the ones you need. All can confuse the AI.
Additionally, add the name of the Process Execution Parameter Name within the Customize Post-Action Instructions if you wish to return it back to the user.
Please extract the SystemMessage from the process output and return it like this: SystemMessage : <Value>
Behavior Configurations
If the Response Directions are too long, loading the AI Agent and starting the conversation can take too long, try to keep it short for context and purpose only. If that does not work, leave it blank.