Sandbox
An isolated environment the agent runs inside: container, VM, or restricted shell. It limits the blast radius.
A sandbox is an isolated environment the agent runs inside: a container, a virtual machine, or a restricted shell. It is the practical answer to the fact that an agent with file and shell access can do real damage very quickly.
The idea is containment. The agent gets a working filesystem it believes is the real one, but writes land in a layer that can be thrown away, network access can be denied or limited to specific hosts, and processes that run out of scope die with the sandbox. A command that deletes a directory deletes the copy.
Sandboxing does not replace permission checks. A sandbox is a boundary that contains accidents and mistakes; it is a weaker guarantee against a deliberate, well-directed agent behaviour, and a sandbox that can reach the network can still exfiltrate what it reads. The two mechanisms cover different things, and a serious setup uses both.
Related terms
Environment
The world the agent acts on: anything outside the harness that it perceives via tool results and changes via tool calls.
Filesystem
A tree of files and directories the agent reads from, writes to, and executes within.
Permissions
The rules that decide what the agent is allowed to touch.
Guardrails
Checks that stop or constrain an action before it happens.
Agent Mode
A preset that bundles permissions and behaviour, like a profile for the agent.
Want this working inside your team?
It starts with one conversation about your workflow.