Sandbox Execution
Running an agent’s code or actions in a walled-off space where mistakes can’t hurt anything real.
Think of it like
A test kitchen sealed off from the restaurant — burn a dish and no customer ever knows.
Example
The agent’s generated script runs in a throwaway container with no network and no access to real files, so a bad command just breaks the sandbox.
How it actually works
A sandbox isolates execution — limiting file access, network, and system calls — so that running model-written code, which you can’t fully trust, stays contained. It’s a foundational safety measure for any agent that executes code or commands, because the whole point is to let it try things freely without letting a mistake or a malicious injection escape. The isolation is never perfect, so sandboxes are layered with other limits.
For product teams
What makes it safe to let an agent run code it wrote itself.
For engineers
Executing untrusted code in an isolated environment with restricted filesystem, network, and syscall access.
Related
- Code Interpreter — The tool that most needs sandboxing.
- Environment — The contained world it creates.
- SWE Agent — Its role in proving software fixes safely.
Read anything AI without the jargon
Look up any term in plain English, or save terms as you read with the free Chrome extension.
Open DecoderAdd to Chrome