Skip to content

Configure an Agent

Give an Agent a clear purpose, durable instructions, and a model that fits its work.

Start with the job the Agent should perform for its users. A good configuration is specific enough to guide choices but broad enough to work across the delivery points where the Agent may appear.

  1. 1

    Name the job

    Use a stable capability or responsibility, not a temporary campaign or channel name.

  2. 2

    Write the instructions

    Explain purpose, expected behavior, boundaries, and when the Agent should ask for clarification.

  3. 3

    Choose the model

    Select from the models allowed by organization policy and appropriate for the task.

  4. 4

    Add governed context

    Attach only the Areas and Connectors the Agent needs — Standards reach it through an Area.

Instructions should make decisions

Avoid long identity prose that does not change behavior. Tell the Agent which users it serves, what a successful answer contains, what it must not assume, and how it should handle missing information.

Keep delivery details elsewhere

Appearance, allowed domains, and installation belong to the widget. The Agent can describe answer style, but it should not contain a domain name or an instruction that is true in only one place.

Choose a governed model

The model picker reflects what the organization allows. A model can be available to view yet unavailable for a particular Agent because of policy, capability, or lifecycle state.

Save a coherent first version

Add the smallest useful context and test it. Large undifferentiated scopes make it harder to understand why an answer succeeded or failed.

A reliable first configuration

Name the Agent after the job it performs, not the channel where it will appear. Write a short purpose, the audience it serves, the tone it should use, and what it must do when evidence is missing. Select an allowed model with the capabilities the job needs. Then attach one coherent Knowledge collection or Area and only the Connectors required for the first test.

Run representative questions and inspect the evidence. Add scope when a real gap proves it is needed. If the Agent will answer in more than one place, keep place-specific instructions out of it and configure them on the widget instead.