Skip to main content
A workspace is where an agent’s execution state lives: a persistent volume plus a session-continuity key. The VM that runs on it is ephemeral — the workspace volume (files, cloned repos, conversation transcript) is what carries everything forward between runs, chats, and SSH sessions.

​
Keyed vs anonymous

Set the keyed-workspace TTL per agent:
When a workspace expires, its volume is reclaimed with it. Continuing or running on a workspace refreshes the clock.

​
Concurrency

One live session per workspace: runs on the same session key queue and execute in order, while runs on different workspaces execute in parallel. This makes a session key a safe serialization point — a schedule that updates the same report every morning can never trample itself.

​
Repos

There is no built-in repo concept. A repo is whatever the prompt (or a skill) clones onto the workspace volume. gh is pre-authenticated in every VM, so “clone acme/api and open a PR” just works — and the clone persists on the volume for the next run with the same key.

​
Files

  • Browse and download workspace files via the dashboard, or via the API (GET /v1/agents/{agent}/workspaces/{key}/files/...) while a VM is up.
  • Run deliverables — files the agent stages in ~/outputs/ are snapshotted onto the run when it finalizes, so they remain available after the VM is gone.

​
VM lifecycle

  • A VM boots when a run, chat, or SSH session starts on the workspace and is destroyed when the execution finishes. The volume always survives.
  • Idle workspace VMs are suspended (Firecracker snapshot — no compute) and woken on message, so a live session resumes exactly where it was, mid-turn if necessary.
  • An execution’s teardown never kills a VM you are SSH-ed into.