Skip to main content
The agent’s computer runs Linux, and the agent has a full shell on it. This is how it runs code, installs what it needs, and drives command-line tools. You never have to ask for this in shell terms. Describe the outcome and the agent picks the commands.

​
What it uses the shell for

  • Running code. Python, Node, whatever the job needs, on real data rather than in a description of the work.
  • Installing tools. It adds packages when a task calls for them.
  • Processing files. Converting, extracting, transforming, and compressing what is in files.
  • Driving CLIs. Any tool that ships a command-line interface, including git and gh.
  • Long jobs. Builds and batch work run in the background while the conversation continues.

​
Opening it yourself

The Computer panel has a Terminal tab: a live shell into the agent’s sandbox. It is useful for inspecting what the agent built, running a quick command, or debugging without asking the agent to do it for you. The terminal is only shown when you have edit access on that agent.
You can also reach the sandbox from your own machine. In the Skydive CLI, /sandbox opens an interactive terminal, and /sandbox <command> runs one command and streams the output.

​
What persists

The sandbox is ephemeral. It sleeps when idle and is rebuilt when it wakes, so anything installed at runtime is gone on the next boot. Work that has to survive goes in one of three places: For a package the agent needs on every boot, it adds the install line to its own Dockerfile in the repo, so the next sandbox is built with it.

Apps & webserver

Host what it builds at a URL you can open.

Sizes

When a job needs a bigger machine.