Agent Script Pattern: Reference Resources Directly in Reasoning Instructions

Reference subagents, actions, and variables directly in the prompt text of your reasoning instructions. Use @ mentions and the {curly brace} syntax to reference variables and reasoning actions directly.

Why Use This Pattern

Reference subagents, actions, and variables directly in your reasoning instructions using special syntax. Use @ mentions and the curly brace syntax to reference variables and actions directly in your prompt text.

Pattern Example: Display a personalized order summary by referencing variables like order name, date, and status, and guide the user to the lookup_order action or Returns subagent when they need more help.

Reference Syntax

Subagents: {!@subagents.<subagent_name>}
Actions: {!@actions.<action_name>}
Variables: {!@variables.<variable_name>}

Reference Variables in Output

Reference variables directly to include specific data in the response.

1reasoning:
2  instructions: ->
3    | Refer to the user by preferred name {!@variables.preferred_name}.
4      Output a summary to the user in the following format:
5      Order Name: {!@variables.order_name}
6      Order Date: {!@variables.order_date}
7      Order Status: {!@variables.order_status}

Reference Actions in Instructions

Reference actions directly to tell the LLM which reasoning action to use.

1reasoning:
2  instructions: ->
3    | If the user wants information about a past order, ask for the Order ID
4      or Restaurant Name and use {!@actions.lookup_order}.
5      If the user asks for a service agent or seems upset,
6      go to {!@actions.go_to_escalation}.
7      If the user wants to make a return, confirm their order ID and
8      call {!@actions.create_return}.

Combine Variable and Action References

Use both types of references together for complete, specific instructions.

1reasoning:
2  instructions: ->
3    | Refer to the user by their name {!@variables.member_name}.
4      Show the user their current order summary: {!@variables.order_summary}
5      at the start of the conversation or if they specifically request it again.
6      If the user wants to make a return, confirm their order ID and
7      call {!@actions.create_return}. If returns are not eligible
8      ({!@variables.order_return_eligible} is False), explain that
9      the order is not eligible for return.

Conditional Output with References

Use references inside conditional prompts as well.

1reasoning:
2  instructions: ->
3    if @variables.loyalty_tier == "Gold":
4      | Thank the customer for being a Gold member.
5        Their current points balance is {!@variables.points_balance}.
6
7    if @variables.loyalty_tier == "Platinum VIP":
8      | Welcome back, valued Platinum VIP member {!@variables.member_name}!
9        You have {!@variables.points_balance} points available.

Tips

Reference resources when clarity is required: In general, the agent can figure out which resource to use. In cases where you have many actions, for example, add a reference to help the agent pick the right one. Direct references are a stronger signal to the LLM, so it’s more likely to select the right variable, action, or subagent.