Skip to main content

AI Teammates and Mailbox Agents

Outermind helps you design an AI teammate by starting with the work you want handled. You do not need to know which agent type to choose before you begin.

Start with the work​

Go to Workforce > Agents and select Create agent. The design interview asks a few focused questions about:

  • What should trigger the work
  • What a good result looks like
  • Which systems and knowledge sources are needed
  • What the agent may do on its own
  • When and to whom it should escalate

You can type your answers or use Live Audio when it is available. The interview produces a reviewable design before anything is created.

The examples on the first page are conversation starters, not fixed templates. You can edit the example text, start with your own idea, or ask the interview to help you figure it out.

Mailbox agents​

A mailbox agent monitors one shared mailbox and one or more folders in that mailbox. Inbox is the usual starting folder, but a different folder can be selected when work is intentionally routed there.

Use a mailbox agent when the work begins because a message arrives. Common examples include:

  • Customer or employee support requests
  • Accounts payable invoices and vendor questions
  • Sales inquiries
  • Compliance or document intake
  • Requests routed into a dedicated team folder

A mailbox agent can inspect the message, use its assigned tools and knowledge, respond within its approved boundaries, or escalate to a person.

Mailbox agents are not callable specialists. Other AI teammates cannot invoke or directly message them as SMEs.

Specialist agents​

A Specialist Agent, or SME, has no monitored mailbox. It performs focused work when an eligible AI teammate invokes it. Use an SME when expertise should be reusable across AICOS, AIPSO, Personal Assistants, or other permitted SMEs.

For details, see Specialist Agents.

How the interview chooses​

The interview uses your description as evidence, not as an irreversible selection.

  • Work that begins with incoming email usually points to a mailbox agent.
  • Work that begins when another AI teammate asks for expertise usually points to an SME.
  • If your idea contains both patterns, the interview explains the difference and may recommend creating two cooperating agents, one at a time.
  • If the work belongs to AICOS, AIPSO, or a Personal Assistant, the interview explains that role and directs you to the correct setup experience.

You can correct the recommendation during the interview. The review updates to reflect the newly required decisions.

Review before creation​

The review summarizes the design in plain language:

  • Name and purpose
  • How work arrives
  • Expected outcomes
  • Mailbox and folders, for a mailbox agent
  • Invocation access, for an SME
  • Tools and knowledge
  • Supervisor and escalation boundaries
  • Activation choice

Open Advanced settings only when you need to change execution limits, model overrides, tool details, sender restrictions, or other technical controls. Defaults remain visible in the review even when the advanced section is collapsed.

Changing a deployment-affecting setting invalidates the earlier approval. Review the updated design before creating the agent.

Plans, prerequisites, and limits​

The interview can explain any agent role, even when the current subscription does not include it. It never recommends creating something the tenant cannot deploy.

Locked capabilities show the required current tier from the feature matrix and provide an upgrade path. Limits and prerequisites are checked again immediately before creation. Typical prerequisites include:

  • A shared mailbox that is available and not assigned elsewhere
  • Microsoft 365 application authorization
  • Required tools and their application connections
  • Configured knowledge sources
  • Available agent capacity

You can complete supported prerequisites and return to the same interview without losing the design.

Creation and activation​

Outermind creates the agent inactive, applies the approved design, and verifies every required resource. The agent is activated only after the required checks succeed and activation was requested.

If a resource is still provisioning, the page identifies the exact item that needs attention. Successfully created resources are preserved, and retry continues from the incomplete step rather than starting over.

Managing an existing agent​

Open Workforce > Agents, select an agent, and choose Edit to adjust its configuration. Editing an existing agent remains separate from the design interview.

Use the agent detail and activity pages to:

  • Pause or resume the agent
  • Review execution history and errors
  • Change assigned tools and knowledge access
  • Adjust sender restrictions or escalation settings
  • Test behavior before relying on it for production work

Testing agents with chat​

Use the agent's test chat to try representative requests before production use. Confirm that the agent follows its boundaries, uses only expected tools, cites the right knowledge, and escalates when required. Test chat does not replace a mailbox delivery test for mailbox agents.

Step 6.7: Configure Long-Running Tasks (Advanced)​

Long-running controls are available in Advanced settings and when editing an existing agent. Increase execution time only when the work genuinely needs it. SME research and analysis may need a longer window; routine mailbox triage should remain bounded so one message cannot consume excessive resources.