Define the job.
Describe goals, tone, boundaries and when the agent should move the conversation elsewhere.
Build an agent from one screen: define its instructions, connect a model, choose knowledge and skills, control assignments, then test the full behavior before it meets a customer.
Create an agent ↗The setup follows the same order as the product: identity, instructions, model, skills, tools, limits and a live test beside the form.
Describe goals, tone, boundaries and when the agent should move the conversation elsewhere.
Select an enabled model provider connection and enter the exact model ID.
Attach enabled knowledge bases and add instructions for when the agent should search them.
Allow only selected standard or custom fields, then describe when each should be updated.
Choose allowed agents, users and teams. Mentions can add a target directly from the instruction.
Bind approved integrations, connections and tool-specific instructions without exposing every action.
Agent instructions can mention allowed assignment targets, knowledge bases and customer fields. The reference is stored as structured data while remaining readable inside the editor.
Refer to an allowed destination and let the assignment skill use it.
Point the agent at the right approved source for a task.
Insert a field or its current value into the skill instructions.
The test panel runs the same configuration with a sample customer record. Activity shows model requests, tool calls, assignments, contact updates, sources and failures.
Test customer questions while preserving the preview conversation context.
Provide realistic input and see the contact record change after a skill runs.
Review which connection and tool were used and the outcome of every call.