Skip to main content

Configuring Specialist Agents

Use the agent design interview to create a new Specialist Agent. This page explains how to review and change an SME after it exists.

Open an existing SME​

  1. Go to Workforce > Specialist Agents.
  2. Select the SME or choose Edit.
  3. On the shared agent admin page, choose the section you need from the left menu. Available sections depend on your permissions.

SMEs now open on Overview at /workforce/agents/<agent-id>, the same destination used by other agent kinds. Command palette results open this page too. The configuration link from an SME's learned-patterns page opens its Knowledge section.

Older SME edit bookmarks and notification links redirect to the shared page. If their target section is unavailable, the page opens Overview.

Identity and discovery​

The friendly name helps people recognize the SME. The discovery description helps AI teammates decide when the SME is the right expert to invoke.

A strong discovery description states:

  • The domain the SME covers
  • Representative requests it handles well
  • The result it returns
  • Important limits or refusal boundaries

Keep the description specific. A broad label such as "business expert" makes routing less reliable.

Tools and knowledge​

Grant only the tools and knowledge sources required for the SME's approved work.

  • Tool access is tenant-scoped and subject to the current subscription and application authorization.
  • Knowledge sources should be authoritative for the SME's domain.
  • SharePoint and other scoped sources may require per-agent access configuration.
  • A missing prerequisite can make the SME inactive until the configuration is repaired.

Model override​

By default an SME uses your tenant's job assignments from the AI Models page. To pin this SME to a specific model instead, open the model picker from the SME's settings and choose a model, or choose Use default to go back to the tenant default.

Who can invoke the SME​

An agent can invoke an SME once the SME's tool is switched on for that agent. Open the calling agent's Tools settings (for the AI Chief of Staff, Boardroom > Settings > Tools), find the SME under Agent Delegation & Coordination, and turn it on. That one toggle governs everything: the SME's dedicated tool and capability search both read the same grant, so there is no second place to enable access.

To see which agents hold a given SME's tool, open the SME's tool under Capabilities > Tools and review its assigned agents.

Invocation policy​

The invocation policy is an SME-side restriction layered on top of the tool grant. New SMEs default to Available to eligible AI teammates in this workspace, which adds no restriction beyond the tool toggle. Use exceptions to deny a particular caller, or switch to an explicit caller list so that only named teammates may invoke the SME even when they hold its tool.

Every invocation still checks:

  • The current SME invocation entitlement
  • The caller's tenant, active status, and lifecycle
  • The caller's grant of the SME's tool
  • The SME's tenant, SME flag, active status, and lifecycle
  • The live policy and per-agent exception

An SME can invoke another permitted SME and can reply in an existing direct conversation. It cannot initiate a new direct message. Mailbox agents cannot be invoked or directly messaged as SMEs.

Execution and escalation​

The SME's invocation timeout is no longer configurable from the console. The Execution Limits tab holds the autonomy level, which sets how much the SME can do on its own before a human approval is required, plus its cost and usage limits.

Select a supervisor who can answer questions or approve escalations. Configure clear thresholds for legal, financial, security, or other high-impact decisions.

Safe changes​

Test material changes before relying on them in production. Confirm that:

  1. The SME appears in discovery for representative requests.
  2. An eligible caller can invoke it.
  3. A denied caller cannot invoke it.
  4. The SME returns a useful result to the caller.
  5. Escalations reach the intended supervisor.

Pause the SME while repairing a broken tool, knowledge source, or policy. Resume it only after the required checks pass.