Agent outputs die when the VM sleeps: octomind's design for deliverables that survive

donk8r · reddit · 2026-09-05

octomind's engineering deep-dive on a default failure of agent infrastructure: an agent spends 20 minutes writing a report, the host VM gets reclaimed on its idle timer, and the download link dies because output lives on a machine designed to go away.

The team rebuilt around one publish tool:

The agent declares a path and title; it never decides where files land. The tool description teaches restraint first — "if you are unsure whether something is a deliverable, it is not."
Publishing copies the file, verifies the hash, and unlinks the original (no flag, to avoid double billing). Files land in an account-wide shared directory mounted on every machine; downloads stream from the control plane, so the producing machine can be asleep.
Security: bytes never come from the API origin (agent-authored HTML near a logged-in session is stored XSS), so they're served from an isolated user-content domain where only inert types render inline. Private links are signed and expire in an hour; public downloads are metered, since container egress sampling is blind to control-plane-served files.
Honest caveats: free-plan workspaces are reclaimed after 7 days including artifacts, and publishing means delete really deletes.

The author closes with an open question: should deliverable judgment sit with the agent's restraint clause, or should users name deliverables explicitly?

Original post →

Agent outputs die when the VM sleeps: octomind's design for deliverables that survive

More from coding & agent