Administration & GovernanceAgents & Operations

Manage Agents

View as Markdown

Agents combine an approved model, instructions, knowledge, and Actions into a reusable experience. Admins enable deployment-level capabilities; Admins and Curators configure Agents within their permitted resource scope. People who use an Agent receive only the access shared with them.

Launch order

Treat every Agent as a reviewable production application, not as a temporary prompt bundle.

StageOutputRelease condition
DesignAgent name, target task, target users, and prohibited behavior.The purpose is clear and not overly broad.
ConfigureModel, system instructions, knowledge sources, and Actions.All resources come from approved configuration.
TestRepresentative questions, failure questions, and access-boundary questions.Answer quality is acceptable, and unauthorized knowledge or tools are not used.
PublishSharing targets, owner, and rollback path.Access scope, audit method, and maintenance owner are confirmed.

Agent building blocks

Building blockAdmin or Curator responsibilityUser and runtime boundary
ModelChoose an enabled language model and its deployment configuration.Users can select only the shared Agent; runtime uses its configured model.
InstructionsDefine the task, behavior, and operational limits.Users provide requests within those instructions; runtime applies them to each run.
KnowledgeAttach approved knowledge sources when the deployment makes them available.Users cannot expand a shared Agent’s knowledge scope; runtime retrieves only configured content.
ActionsAttach an approved MCP or OpenAPI Action that is currently available for selection when the editor can edit the Agent.Users authorize any prompted action use; runtime invokes only enabled tools.

Create and configure an Agent

  1. Sign in as an Admin or Curator and open Agents.
  2. Create an Agent, give it a clear name and purpose, then select an enabled model.
  3. Add instructions that describe the intended task and limits.
  4. Save the draft and review its configuration before making it available to others.

The Admin or Curator editing the Agent is responsible for its configuration and for selecting from models already enabled for the deployment. Do not treat a user request as permission to change those settings.

Attach knowledge and Actions

Attach only knowledge sources and Actions that have been configured and approved for the deployment. Admins can manage and attach any available Action across the deployment. A Curator can manage Agents, connectors, and document sets for groups for which they are a curator; a Global Curator has the same group-resource permissions for groups of which they are a member.

Action ownership is separate from those group boundaries. Either Curator role can create MCP and OpenAPI Actions but can modify, delete, or maintain protected credentials only for Actions they created; Global Curator status does not expand Action ownership.

A Curator can attach any currently available and selectable Action to an Agent they can edit. The same Agent-edit boundary applies when attaching a permitted document set or connector-backed knowledge source.

When a person uses the Agent, they can work only within the shared experience and any authorization prompts presented to them. At runtime, AIvis uses the Agent’s saved configuration and invokes only the knowledge or Actions that are enabled for that run.

Publish and share

Publish an Agent only after its configuration has been reviewed. Share it with the intended people or groups according to the deployment’s access controls, and remove access when it is no longer needed.

Publishing an Agent makes its approved configuration available; it does not grant administrators’ configuration permissions or unrestricted access to every connected service.

Verify access

As an authorized Admin or Curator, test the published Agent with a representative approved request. Confirm that the intended people can find and use it, that people outside the shared scope cannot, and that attached Actions behave as expected when authorization is required.

For runtime verification, review the result of a real test run and any reported tool error. If a model, knowledge source, or Action is unavailable, correct the relevant configuration or authorization boundary before republishing.