Agents

System instructions#

System instructions serve as the primary set of rules governing your agent's behavior. They define:

• Persona & Tone: Instructing the agent to speak professionally, concisely, or with a helpful customer support demeanor.
• Scope & Boundaries: Setting guidelines on what topics the agent is authorized to speak on, and how to handle off-topic inquiries.
• Workflow Guidelines: Providing specific procedures, such as verifying order details before discussing refund options.

Starter suggestions#

Starter suggestions guide users to launch conversations. These appear as clickable prompt chips in the web widget when users first open it. Examples include:

• "How do I rotate API keys?"
• "What is your return policy?"
• "Track my order"

Suggestions reduce friction and guide users to common, automated answers.

Playground testing#

Playground testing allows simulating customer questions and verifying responses in a test environment. During testing, you can:

• Verify which sources were retrieved and referenced for the answers.
• Refine guidelines and instructions based on real-time feedback.
• Inspect agent decisions step-by-step.

Conversation trace#

To understand how your agent operates and verify accuracy, trained.chat provides a detailed conversation trace. For every message processed, you can inspect:

• User Question: The query analyzed by the agent to understand intent.
• Knowledge Sources Used: The exact fragments of documents and pages retrieved from the knowledge base.
• Actions Executed: The APIs triggered, parameters parsed, and responses returned.
• Response Time: The latency profile of each step in the agent's decision loop.