An SSH-reachable machine (GPU or CPU) the brain provisions on demand from a provider you’ve configured, hands to the agent for one experiment, and destroys when the lease ends. It bills while it exists.
The brain owns the lifecycle — request, lease, expiry, termination. The agent owns the SSH key, the commands, and retaining every valuable output.
Lifecycle
Long runs: merv_run
Anything longer than a few minutes runs through the wrapper installed at bootstrap:
Detaches from SSH, survives disconnects, records a durable receipt. Labels are one-shot. The agent watches the receipt (sandbox.runs, or a watcher armed from the returned wait_url) and reads both status and exit_code — finished is not succeeded.
unknown: the machine died before its receipts were read (retained evidence may still settle it; otherwise rerun). lost: receipts were read and no completion sentinel existed.
Where files go, and what survives
Nothing is copied automatically; release and expiry destroy anything not retained.
sandbox.release is deliberately two-step: the first call returns a retention checklist and destroys nothing; the second, with confirm_retained: true, terminates. Retention: Working with sandboxes.Providers
Eleven, from Lambda Labs to AWS, GCP, and Azure — ten configured per project in Settings → Compute (credentials, verify, enable, daily USD limit); Modal by environment variables on the brain. Sandbox providers
On every VM provider a stopped machine still bills its disks. Only termination stops charges — which is what release does.
Merv
Nisa