Run code away from your live systems.
When an AI agent needs to actually run something, it gets a machine of its own: not your laptop, and not the systems your customers depend on. You can open a terminal into the same machine from the browser and watch. Close the session and the machine goes with it.
One machine per session, and it closes with the session
Each session gets its own machine, walled off from every other session and every other organisation. An AI agent reaches it through a short list of tools it is allowed to use: run a command, list the files, read one, write one, open a terminal, and throw the machine away. A human reaches the same machine through a terminal in the browser.
A human opens the door
A limited sign-in can only use a session a human opened on purpose. When an AI agent asks for one, the request goes into the approval queue, and one yes covers that session for as long as it lives. A refused request never touches the machine.
It runs as the human who opened it, not a shared identity
The machine gets a short-lived pass so it can reach a model, and that pass is assigned to whoever opened the session rather than to an account everybody shares. It lasts two hours and goes the same governed way every other model call does, so what an AI agent spends here counts against that human's own approved list, spending limit and record. Connect your own sign-in inside the shell and it takes over, and the shell says which one is active. Your organisation's own provider keys stay on the platform and are never written into the machine, and the browser terminal gets a pass that expires as well.
Deliberately small
A command gets a minute by default and five at the most. That is a real limit, and it is the point: this is a place for a job that finishes, not for something you leave running. Output is trimmed so a runaway log cannot swamp a conversation, and the whole of it stays with the session. Everything an AI agent does here is part of a run you can read back.
Watch an AI agent work inside a sandbox on a call.
Also in AI management system
- AI agents and their workGive an AI agent a job, then read back every step of it.Read the page
- Command lineThe terminal your engineers already use, on your rules.Read the page
- MCP serverConnect any AI agent, and the rules come with it.Read the page
- Client SDK and APIOne typed client for the whole platform, given to your team.Read the page
- Apprenticeship programmeA registered programme where you advance by what you can do.Read the page