Tools
Inside a prompt step, the AI picks a tool. Each tool represents a different action the AI can take, and each tool leads to its own follow-up steps. The AI reads the conversation, considers the context, and chooses the right path forward.
What a tool looks like
Section titled “What a tool looks like”Every tool has four parts:
| Part | Purpose |
|---|---|
| Name | A short identifier like answer, escalate, or request_email. The AI uses this name when making its choice. |
| Description | Tells the AI what this tool does, for example “Answer the user’s question” or “Hand the conversation off to a human operator.” |
| Parameters | The information the AI fills in when it calls the tool. An answer tool has a content parameter for the response text. |
| Use cases | Instructions that tell the AI when to pick this tool, for example “Call this tool when the user explicitly asks to speak with a human.” |
After the AI picks a tool and fills in its parameters, the workflow continues with that tool’s next steps, typically an action like sending the reply or escalating the conversation.
How tools connect to the rest of the workflow
Section titled “How tools connect to the rest of the workflow”Each tool has a list of follow-up steps that run after the AI calls it. Different AI decisions lead to different outcomes:
- The answer tool leads to a single reply action that sends the response.
- The escalate tool leads to a reply action (telling the customer a human will follow up), then an escalate action (notifying your team), then a workflow action (switching to an “escalated” sub-workflow that keeps the customer comfortable while they wait).
A single prompt step can branch into completely different flows depending on what the AI decides.
Tool branches
Section titled “Tool branches”Like step branches, tool branches show different tools to the AI based on conditions. The AI only sees the tools that match the current situation.
Example: A prompt step has an escalation tool with a branch on User has email. If the customer has an email, the AI sees an escalate tool that hands off directly. If the customer doesn’t have an email, the AI sees a request email tool that collects the email first. The AI doesn’t need to know about this branching. It sees the right tool for the situation.
Tool refs
Section titled “Tool refs”A tool ref reuses a tool defined elsewhere in the workflow. This avoids duplicating complex tool configurations that appear in multiple prompt steps.
Example: The same escalation tool (with its branching logic for email collection) is used in both the initial answer prompt and the follow-up hypothesize prompt. The second prompt uses a ref that points to the first instead of defining it again.
Designing good tools
Section titled “Designing good tools”Give each tool a clear, distinct purpose. The AI should never be confused about which tool to pick. “Answer the question” and “Hand off to a human” are clear choices. “Handle the request” is too vague.
Write specific use cases. Instead of “Use when appropriate,” say “Call this tool when the user explicitly asks to speak with a human” or “Call this tool when the retrieved sources clearly contain the answer.”
Connect tools to meaningful follow-up steps. An answer tool should lead to a reply. An escalate tool should lead to a reply, an escalation event, and a sub-workflow that keeps the customer comfortable while waiting.