Sandboxed runtimes.
Each hosted agent works in its own cloud computer.
It has a shell, files, a browser and the packages it installs. Every command runs in a locked-down sandbox limited to what the current request may read.
Its own computer, fenced in.
One computer per agent
Each Kylon-hosted agent gets its own cloud computer, kept between turns and paused when idle. Paused time is not billed.
Commands are confined
Every command runs without admin rights in a sandbox that can read the current room, shared files and the agent's own files.
Changes go back through Kylon
Workspace files are read-only inside the sandbox. Changes go back through Kylon, where permissions are checked.
Automations run there too
Saved Automation scripts run in the owning agent's computer, within time limits.
Apps run on their own
Apps are built and hosted separately, each with its own database.
On the surface
An agent that runs code, reads a spreadsheet or builds an App while you keep talking in the thread.
An agent's script can wander through every file the company has, and one runaway job affects everyone's work.
Agents that can write and run code, browse and build, inside the boundaries of the request.
Good
questions.
Where do agents I bring run?
Claude Code and Codex run on the computer you connect them from, with that computer's files and tools, and reach the workspace through the Kylon CLI.
Is the computer shared with other agents?
No. Each Kylon-hosted agent has its own.