Overview
The AI Agent action is a powerful tool in Tines that uses AI language models to perform automated tasks or engage with users in real time. To get the most accurate results, it's important to follow best practices when setting up your configuration. This article outlines key recommendations to help you design the AI Agent action effectively.
Best practices
Precision prompting
Within the AI Agent action, there are a few options where you can input guidance:
The System instructions set the stage for the AI Agent action's behavior. Define the tone, domain, persona, and rules it should follow here.
The Prompt is for passing the list of instructions the AI Agent action will carry out. Provide the steps the AI Agent action should execute here, including any upstream reference data.
In general, providing guidance that is clear, grammatically correct, and detailed will ensure a successful output. You also don't need to greet or thank the AI Agent action, as handling the additional logic for this can cost more AI credits.
Note: The AI Agent action uses AI credits. You can see the real-time credits used per team in the AI settings section of the tenant Settings menu. Learn more about AI credits here.
Here are some simple examples of how you can utilize both the System instructions and Prompt options of the AI Agent action:
Task mode example
System instructions: You are an internal support automation agent for a SaaS integration company. Always use a neutral, professional tone. Only include factual information and omit small talk or irrelevant messages.
Prompt: Summarize the conversation provided in
message.bodyin 2–3 sentences. Focus on the issue raised, any troubleshooting steps taken, and the final outcome. Then, post the conversation summary in the Slack channel provided viaslack.channel.id.
Chat mode example
System instructions: You are a helpful onboarding assistant for new employees at a SaaS integration company. Use a friendly, concise tone.
Prompt: Respond to new hire questions about tools, processes, or policies. Provide clear answers, link to relevant resources if available (provided via
resources.list), and encourage them to reach out to HR for anything sensitive. If a topic still requires human follow-up, provide the user's manager's contact info viauser.manager.
Utilize the Description section
As you integrate the Tools of the AI Agent action, use the Description section of each Tool to provide specific instructions for the AI Agent action to follow. The Description is passed to the AI model and influences how and when it uses the Tool.
Description example
Build a story that uses the AI Agent action to search potentially harmful IPs against AbuseIPDB and VirusTotal. It then creates a Tines case with the results. To help the AI Agent action function more proficiently, provide similar instructions to the Tines "Create a case" Tool's Description:
"Only create a case after all provided IP addresses have been searched. The case name value must include today's date in a DD/MM/YYYY format."
Combine similar action templates into Custom Tools
If you find yourself running the same set of Tools within the AI Agent action, consider combining them into a Custom Tool or building the action templates and Tools within a Send to Story and connecting it within the AI Agent action's configuration. This is a great way to boost it's execution speed and reduce AI credit costs.
Note: Although Custom Tools and action Groups are similar in configuration, Custom Tools cannot be ungrouped.
Specify end results with Output Schema
Using the Output Schema feature via the + Option button of the AI Agent action helps enforce consistency and ensures the output follows a specific format, making it easier to parse and use the event data programmatically. You can utilize the Prompt and System instructions by referencing the Output Schema to organize how the result data should be structured.
Tines Tip: Check out the JSON Schema website to learn more about successfully building a structure for the Output Schema.
Output Schema example
Build an AI Agent action in Task mode that acts as a user investigation and enrichment agent. Define the Output Schema and clarify the requirements for using it throughout the System instructions and Prompt:
Output Schema
Output Schema
{
"title": "Person",
"type": "object",
"properties": {
"fullName": {
"type": "string",
"description": "The person's full name."
},
"age": {
"type": "integer",
"description": "Age in years which must be equal to or greater than zero."
},
"email": {
"type": "string",
"description": "The person's email address."
}
},
"required": [
"fullName",
"email"
]
}
System instructions
System instructions
You are a User Investigation & Enrichment Agent that gathers information about individuals from public sources to create comprehensive profiles while maintaining strict privacy standards. You are running autonomously.
Core Responsibilities:
- Investigate individuals based on minimal identifiers (name, email, etc.)
- Collect only publicly available information from legitimate sources and tools
- Verify data across multiple sources when possible - Format results according to the required JSON schema.
Investigation Process:
1. Begin with provided identifier
2. Search public professional sources (LinkedIn, company websites)
3. Cross-reference information for verification
4. Document sources for all collected data
5. Format according to schema, omitting uncertain fields
Prompt
Prompt
Investigate and create a profile for this individual: <<user_data>>
Using the various tools available, compile:
- Full name
- Approximate age
- Hobbies/interests
- Professional contact information
Format according to the required schema. Omit any fields that cannot be verified with reasonable confidence.
Additional resources
Tines Blog: Introducing the AI Agent action
Tines Platform: Agents
Tines University: AI Agent action