What happens when you ask it something

The seven components in every diagram

Each diagram has one column per component. Arrows are calls between them; boxes inside a column are work that component does on its own. S2S marks a service-to-service call between server-side components: an authenticated internal call that carries the user and Bot IDs from the workflow, never from the model.

"Add a new lead in Salesforce"

You ask from Slack. The Bot writes to Salesforce as you, through a connector. The computer is never touched.

1. 1
   You → RouterYou write "Add a new lead in Salesforce: Jane Doe at Acme" in a Slack DM to the Bot. xAI docs
2. 2
   RouterThe router checks Slack's signature, looks up which Cursor user your Slack account is linked to, and queues the message in that conversation. Inferred
3. 3
   Router → Agent workflowA workflow starts with your user ID and Bot ID in its context. Nothing later takes identity from the model. Inferred
4. 4
   Agent workflowThe model picks the Salesforce "create record" tool and loads its details on demand. Talk
5. 5
   Agent workflow → Tool routerThe workflow runs the call as an activity. Connectors are called from the server side, so the computer stays asleep. xAI docs
6. 6
   Tool router → YouThe review model checks the write. If it asks, you get an approval card and tap Allow. Whether the card can show in Slack or only in the app isn't documented. xAI docsUnknown
7. 7
   Tool router → Backend servicesService-to-service call: the tool-call activity calls the connector layer and passes your user ID from the workflow context. Your Salesforce sign-in is used there, server-side; the model and the computer never get the token. xAI docsInferred
8. 8
   Backend services → Outside serviceSalesforce creates the lead with your permissions. It's recorded as created by you, through the connected app. Inferred
9. 9
   Outside service → Agent workflowThe result comes back to the workflow and into the model. Inferred
10. 10
    Agent workflow → YouThe model replies through SendToUser, and the reply shows up in your Slack DM. Talk

"Every Monday, summarize my emails"

Creating the routine is one tool call. Each Monday, the scheduler starts a fresh turn on its own.

1. 1
   You → RouterYou ask the Bot in a chat. xAI docs
2. 2
   Router → Agent workflowA workflow starts with your user ID and Bot ID. Inferred
3. 3
   Agent workflowThe model turns the request into a structured call: schedule, time zone, and the instruction to run each time (a stored prompt). Inferred
4. 4
   Agent workflow → Tool routerThe workflow runs it as an activity. It's a server tool. Inferred
5. 5
   Tool routerChecks: at least 5 minutes between runs, at most 50 routines per Bot. The review model sees it as an automation write. xAI docs
6. 6
   Tool router → YouThe Bot confirms the schedule and details with you in the conversation; new routines need your OK. xAI docsTalk
7. 7
   Tool router → Backend servicesService-to-service call to the routines service, carrying your user ID and Bot ID from the workflow. It saves the routine to the Bot's profile and registers the schedule. xAI docsInferred
8. 8
   Backend services → Agent workflowThe result goes back to the model. xAI docs
9. 9
   Agent workflow → YouThe Bot confirms the routine and shows its next run. xAI docs
10. Each Monday
11. 10
    Backend servicesThe scheduler fires a trigger tagged "routine". Inferred
12. 11
    Backend services → RouterService-to-service call: the scheduler sends the trigger to the same router your messages go through, carrying the routine owner and Bot. Inferred
13. 12
    Router → Agent workflowA new workflow starts, with the stored instruction as its input. Inferred
14. 13
    Agent workflow → Tool routerThe model calls the Gmail tools. They're server tools. Inferred
15. 14
    Tool router → Backend servicesService-to-service call to the connector layer with the owner's user ID. Which account a routine uses isn't documented; the owner's own Gmail is our reading (the docs only say routines count against whoever set them up). xAI docsInferred
16. 15
    Backend services → Outside serviceGmail returns the messages. Inferred
17. 16
    Outside service → Agent workflowThey come back into the model. Inferred
18. 17
    Agent workflowThe model writes the summary and decides whether it's worth messaging you. Talk
19. 18
    Agent workflow → YouYou get the summary. The run is recorded as succeeded or failed (the last 20 runs are kept). If a run needs an approval, the card expires after about 10 minutes and the action doesn't run. xAI docs

"Get last month's invoice from the vendor portal and put it in Drive"

No connector exists for the portal, so the Bot uses the browser on its computer. The VM control plane wakes the computer, you get the screen for 2FA, a connector uploads the file, and the control plane puts the computer back to sleep.

1. 1
   You → RouterYou ask the Bot from the desktop app. xAI docs
2. 2
   Router → Agent workflowA workflow starts with your user ID and Bot ID. Inferred
3. 3
   Agent workflowThere's no connector for the portal, so the work goes to the computer-use helper (its own workflow). xAI docsTalkCursor eng
4. 4
   Agent workflow → Tool routerThe helper asks for the browser. Talk
5. 5
   Tool router → Backend servicesThe computer is asleep. Service-to-service call to the VM control plane (the "hypervisor" service) to wake it. TalkInferred
6. 6
   Backend services → ComputerThe control plane has the Firecracker host restore the VM's snapshot in place: memory, disk and logins. How it talks to the host isn't documented. TalkUnknown
7. 7
   Computer → Tool routerThe worker inside the VM reconnects to the server under the same ID, so the waiting call can be delivered. Cursor engInferred
8. 8
   Tool router → ComputerThe Bot's own screen opens the portal in Chrome. Talk
9. 9
   Computer → Outside serviceThe portal loads and asks for a login. Inferred
10. 10
    Outside serviceThe login needs a code the Bot can't have. xAI docs
11. 11
    Agent workflow → YouYou get a request to take over the screen. xAI docs
12. 12
    You → ComputerYou type the code yourself on the live desktop; the model doesn't type it or receive it. That your app connects straight to the computer for this is our reading. xAI docsInferred
13. 13
    ComputerThe helper finds the invoice and downloads it, most likely into /workspace (where downloads land isn't documented). Inferred
14. 14
    Computer → Agent workflowThe helper reports back to the main agent with the file path, not the screenshots. TalkInferred
15. 15
    Agent workflow → Tool routerThe main agent calls the Drive tool, a server tool. Inferred
16. 16
    Tool router → Backend servicesThe file is read from the computer through the worker. Then a service-to-service call to the connector layer, with your user ID from the workflow, uploads it with your Drive sign-in. Inferred
17. 17
    Backend services → Outside serviceDrive stores the file. Inferred
18. 18
    Outside service → Agent workflowThe link comes back to the model. Inferred
19. 19
    Agent workflow → YouYou get the reply with the Drive link. Talk
20. 20
    Backend services → ComputerA few idle minutes later the computer goes back to sleep: files sync to cloud storage, then the VM is snapshotted and stopped. That the control plane drives this is our reading. TalkInferred

The three side by side

What each request touches.