Persistent AI Feature Design: One Identity, One Mention, One Continuation Rule
A persistent AI teammate earns trust when it has one identity, one mention surface, one continuation rule, and visible permissions.
| By | Prompt & Product — Newsroom |
|---|---|
| Filed | 6 October 2026 |
| Read | 3 MIN |

A persistent AI feature can fail before the first model call: it lands as a new tab, a new bot, a new login, and a place where work disappears. Users tolerate that briefly, then stop trusting the output. The fix is an interface contract: one identity, one mention surface, one continuation rule, and visible permissions. The teammate keeps working after the user leaves, but only under a stated rule.
Name the teammate before the model exists
Give the feature a named identity, not a brand mark. A named participant can be assigned, audited, and removed. Autopilot is a cloud agent configured with a name, role, and goal, with separate directory identity, memory, and admin-controlled permissions. Done looks like a person you can find in the address book, with a job title and a scope.
The restructured Copilot app launched on September 25, with Home, Code, and Autopilot as its three tabs. Scout was rebranded as Autopilot in that launch. The structure matters more than the name.
Do not make the agent a generic assistant. If it owns a workflow, give it the role that workflow implies. A procurement agent should read like the role it owns, not a chatbot with a job title.
One mention surface keeps the state model simple
Choose the place where the work already happens, and make the agent mentionable there. Autopilot can be addressed by @mention in Teams, Outlook, chats, channels, and documents. The agent appears where the user is already coordinating, not in a separate console.
One surface means one mental model. If the agent can be mentioned in chat, email, and documents, the user must remember which mention creates which state. That creates confusion. Keep the primary mention surface singular, and treat other surfaces as read-only mirrors or handoff points. The target is one mention that creates the same state no matter which app the user is in.
Write the continuation rule before the first demo
Define what happens after the user leaves. Autopilot continues working when the user is not active. The rule should say who may start it, what it may touch, when it stops, and where the result lands.
The Copilot redesign includes Autopilot, a persistent agent that monitors projects and coordinates with employees. Keep the scope to monitoring and coordination, not open-ended control.
A finished rule reads like a short policy: 'Continue until the shipment issue is resolved, then post a summary to the channel and stop.' If the rule is hidden in settings, the user will not trust the off-screen work.
Avoid making continuation sound like autonomy. Users accept delegated work when the boundary is visible. They do not accept a background process that can keep acting without a stated stop condition.
Visible permissions make off-screen work trustworthy
Show what the agent can see, what it can change, and what it has already done. Agent 365 is the control layer for Autopilot, supplying permissions, audit records, and administrative oversight. The user should see a version of that control layer, not just the admin.
During a September 23 presentation, Microsoft CEO Satya Nadella said agents need identities and observable actions. Trust requires visibility: if the user cannot see the action, the action does not count as trustworthy.
Visible permissions should include the cost of running. Usage-based billing applies to the Code and Autopilot parts of the new Copilot app. If the agent can keep working after the user leaves, the user needs to know what it costs and who is accountable for it.
Microsoft's FinOps updates extend Agent 365's spend oversight to Autopilot, Code, and the managed runtime, with Copilot Studio agents due in October. That gives finance a place to see spend before it becomes a surprise.
Run a scenario where the user closes the laptop. The agent should still be able to act within the stated rule, and it should leave a trace that the user can inspect. In Microsoft's demo, an Autopilot called Dot tracked Black Friday preparation across email, Teams discussions, Dynamics 365 inventory records, and spreadsheets, flagged a shipment issue affecting 18 stores, and reported the issue as resolved.
The finished state includes a permission panel the user can open without asking an admin. It shows scope, spend, recent actions, and a handoff note.