Sandbox
Prerequisite for granting autonomy safely: errors in the sandbox do not leak into production.
In one sentence
An isolated, controlled environment where the agent runs code and actions without risk to the real system. Errors in the sandbox do not leak into production. A prerequisite for granting autonomy safely.
Before
Action with unknown risk
What it does
Runs in an isolated environment
After
Error contained, production intact
Example
The agent that generates the batch runs in an environment where it can read the report and write files, but has no email credential and no write permission for the records system.
If it makes a serious mistake, the damage is a bad file in a folder, and no wrong letter reaches a sponsor’s inbox.
In practice
- Start with what the agent cannot do: production credential, writing to real data, external sending.
- Try to break out of the sandbox on purpose before trusting it.
- If the sandbox needs a “just this once” exception, it no longer isolates anything.
How to make it tangible
A separate environment (container, project, or account) where the agent operates without a production credential.
Translated from Portuguese with AI assistance.