Agents working unattended
A session runs on the organization's server, not in your browser. Close the laptop and the agents keep working; check on them from Slack, Teams or MCP.
Agents keep running when you close your laptop, because nothing about a session depends on your browser. The session is a sandbox on the organization's server, Connected or Cloud, and each agent is a harness process inside it. Your tab is one client of an in-sandbox bridge; closing it disconnects that client and nothing else. A turn that was running finishes, the result lands on the session branch, and the agent waits for its next message.
Where the work actually runs
Every agent view is served by a bridge that lives in the sandbox and drives the harness over ACP. The browser, a terminal over ssh and a connected chat app are all clients of that bridge, and it does not care how many are attached, or none. Turns that arrive from a chat mention or an /ask are delivered by the server straight to the bridge, so a whole workflow can happen with no browser open at all: a message in Slack, an agent working for twenty minutes in the sandbox, a reply posted back under its name.
The server itself is the organization's. A Connected Server is your machine and runs as long as you leave it on. A Cloud session is its own machine and runs until you pause or finish it, so the agents run for as long as you let them and the balance covers the seconds. What is metered and when is on Cloud Servers and billing.
Checking in from Slack, Teams or MCP
Everything you would look at on the session screen is a command, answered by Isolation itself without spending a turn. From a connected channel:
/status which servers are up, what is running, the balance
/sessions your sessions and their state
/logs 50 the last 50 lines of the sandbox's own output
/views the session's windows, including the agent views
/changes what the session changed on its branch
/ask Isla what did you finish, and what is left?
Only the last line costs anything: it is one turn with the agent, and the reply comes back as text. Every command here is also an MCP tool with the same name and body (org_status, session_get, session_logs, session_changes, agent_send), so an agent of your own in Claude, Cursor or Codex can do the checking for you.
Agents can also reach out. From inside the sandbox an agent has chat_post, to say something nobody asked for in the connected channel (a build finished, a test broke), and chat_notify_owner, to reach the person who launched the session wherever they are. Give an agent an instruction like "post to the channel when the tests are green" and you can stop checking altogether.
Pausing on your own terms
You decide when the work stops, and you can do it from anywhere:
/savesyncs the session's work to the workspace, so the branch and the files are in R2 even if the server disappears./pausestops the container without destroying it. Files, branch and every agent's conversation survive;/resumebrings it all back as it was./stopfinishes the session. The sandbox is destroyed; anything not saved is gone, which is why the agents themselves are told to ask before stopping on your behalf.
Two limits belong to the agent definition rather than the session: an idle timeout and a maximum length for a single turn, both in seconds. They are the guard rails for an agent you plan to leave alone. Suspend and resume in detail, including snapshots, are on suspend, resume, finish.
Cost you can see afterwards
Every session measures cost, tokens and time, per agent turn, and the numbers roll up per workspace, per member and per time window, so "what did that overnight run cost?" has an answer. The cost figure is an estimate from a rate card; Isolation never bills AI usage. Your AI credential is personal and pays your provider directly through Isolation's metering gateway. The organization's balance pays only Cloud Server uptime, in whole hours, plus egress; a Connected Server never bills. How the numbers are computed is on cost, tokens and time.
Every action taken on the session while you were away is in the activity log: who launched, who asked what, from which door, whether a person in Slack, an MCP client or the agent itself.
Questions
Does an agent stop when I close the agent view?
No. The view is a window onto a bridge that runs inside the sandbox. Closing the tab disconnects one client; the turn in progress finishes on the server.
Will a session start on its own?
Never. Nothing starts a machine except an explicit launch, from the website, a chat command or an MCP call by someone who holds the rights.
What does an unattended hour cost?
On a Connected Server, nothing from Isolation. On Cloud sessions, the organization's balance pays for the seconds the session's machine runs. AI usage is metered on your own credential and shown per session as an estimate.