Create and configure an AI coach generator

This guide provides a comprehensive walkthrough for creating and configuring AI coach generators in the Agent Assist console. It bridges the gap between conceptual overviews and technical tool integration by detailing the specific steps and logic required.

Create a generator

Follow these steps to create a generator:

1. Sign in to the Agent Assist console and select AI coachunder the Featuressidebar.

   Agent Assist console

2. Click Createto open the configuration pane.

Core settings

Configure the following core settings:

• Generator name: Provide a unique, descriptive identifier (for example, upsell-pixel-watch ).
• Version selection: Choose the most recent stable version. Pay attention to the version lifecycle tooltips:
  • Preview: Allowlist only; not yet generally available.
  • Legacy: Pending deprecation; update soon to avoid service disruption.
  • Deprecated: No longer supported; update immediately.
• Generator-level trigger: Defines the default event for the model to evaluate the conversation (for example, On customer messages).

Overarching guidance

Use the Overarching Guidancesection to set global rules that apply across all instructions within the generator. This helps maintain consistency without repeating logic in every instruction.

• Glossary: Define business-specific terms. For example:

  Account authentication is considered completed if the PIN matches.

• Style and tone: Specify how the agent should sound. For example:

  Always be empathetic and avoid jargon.

• Reasoning process: Tell the model how to think. For example:

  Prioritize security-related instructions over upselling.

For more details or examples about overarching guidance, see Best practices: Overarching guidance .

Configure instructions

Generators are composed of one or more instructions. Each instruction defines a specific scenario and the corresponding guidance for the agent. Add instructions using the following parameters:

Component	Description	Requirement or format
Display title	A name used to manage multiple instructions.	Not visible to the LLM.
Display details	Static content shown to the agent.	Supports Markdown for links and rich text.
Instruction trigger	When this specific instruction triggers.	Overrides generator-level triggers.
Condition	When the instruction is applicable.	Visible to the model (for example, "Customer asks about pricing").
Agent action	What the agent should do or say.	Visible to the model; supports step-by-step logic.
Tools	Automated system actions.	Format: ${tool:tool_name/action}

For detailed guidelines on conditions, actions, and system actions, see Best practices: Instructions .

Suggestion quality best practices

To optimize the quality of the suggestions generated by AI coach, consider the following recommendations:

• Concrete examples: If the model provides generic responses, add "For example" sections in the Agent Action. For example:

  For Galaxy phones, recommend Galaxy Watch.

• Message templates: Use message templates in your actions to control exact wording:

  Since we've been talking about [Topic], I'd like to recommend [Product] because [Reason].

• Suggestion deduplication: Enable Suggestion dedupingin the generator configuration to prevent showing duplicate or highly similar suggestions to the agent repeatedly. You can tune the Similarity threshold(default 0.8 ) to control sensitivity. For details and protobuf fields, see Best practices: Suggestion deduplication .

Test with the simulator

Before deploying your generator to a conversation profile, use the simulator to validate its behavior:

• Type messages as a Customeror Agentto see real-time prompts and triggers.
• Upload JSON conversation files (up to 300 messages) to test complex multi-turn flows.
• Verify that Entity extractioncorrectly identifies parameters (like addresses or account numbers) and passes them to the inputParameters of your tools.