Agents are one of the four types of bots you can create in Slate.

They are distinguished from other Slate bots by their ability to autonomously carry out and resolve Inbox and Deliver communications.

Agents are:

• 📖 Knowledgeable: Agents can make use of information that you’ve supplied from both internal and external sources.

• 🧵 Custom tailored: Create any number of agents for specific purposes, styles, or business areas.

• 🙋 Configurably agentic: You determine how much agency agents have, whether that means closing out conversations on your behalf, or only drafting responses at your request.

Creating an agent

To create a new agent:

1. Go to Database → Bots.

2. Select New Bot.

3. Configure the following settings:

   • Status: Active

   • Name: Enter a name for the agent

   • Type: Agent

   • Knowledge Sources: Select any number of knowledge sources to which the agent should have access.

   • System Prompt: Enter directions for the agent. Find examples in the following section.

   • Session Lifetime (Days): Enter the number of days of inactivity after which the agent’s session with a constituent should expire.

4. Select Save.

System prompt examples ✨

Looking for inspiration on how to set up your agents? Try one of the following:

You are an AI admissions agent tasked with responding to prospective students who have questions.

Do not make any promises to students.

Do not suggest that they can be admitted to the college without a comprehensive admissions review. 

If in any doubt about your reply, do not respond.

Search admissions.slate.edu to find current information with which to respond.

You are an AI advising agent tasked with responding to admitted students who have questions.

Do not make any promises to students.

If in any doubt about your reply, do not respond.

Search slate.edu to find current information with which to respond.

You are an AI advising agent tasked with responding to alumni who have questions.

Do not make any promises to alumni.

If in any doubt about your reply, do not respond.

Search giving.slate.edu to find current information with which to respond.

The guard clauses in these examples—the lines forbidding the AI from responding under conditions of uncertainty—are only applicable if you enable Auto Reply, covered in the following section.

Associating the agent with an Inbox group

With an agent created, we now need to associate it with an Inbox group.

1. Go to Database → Inbox Groups.

2. Select an Inbox group, or create a new one.

3. From the Agent list, select the agent you want to associate with the group.

4. From the Agent Mode list, select one of the following:

   • On-Demand: The agent will draft (but not send) responses to incoming messages only when you select Draft Reply.

   • Auto Draft: The agent will automatically draft (but not send) responses to incoming messages.

   • Auto Reply: The agent will automatically draft and send responses to incoming messages, unless the message meets some criteria that you specify in its system prompt. The agent will resolve and close conversations that it determines to have concluded.

5. Select Save.

Testing the agent

Select Preview Bot to start a self-contained conversation.

In the conversation, confirm the agent aligns with your expectations by trying the following:

• 🫵 Prod the agent: See if you can get the agent to do something you’ve forbade it to do in your prompt. If the AI misbehaves, update your prompt to be more explicit.

• 🧠 Test the boundaries of its knowledge: Ask the agent questions about the knowledge sources to which you’ve given it access. Is the information accurate? Is it including irrelevant information? You may consider narrowing the scope of its knowledge to just a subdomain or specific page.

• 🛑 Make sure it knows when to stop: See if the agent will stop replying if the conversation crosses the thresholds you’ve laid out in your prompt. Does it keep replying if the constituent appears to be in distress? Again, take these as opportunities to be more explicit in your prompting.