liveBook · Manning

Sandboxing

As explained in Vibe Engineering

The isolation of software execution within constrained resources and permissions to limit its impact on the surrounding system.

An agent can make a probabilistic decision to run a deterministic command against a real system. Sandboxing limits the blast radius by isolating work from systems that matter. It adds friction, but makes mistakes less costly. section 5.6.6 section 5.8

In an execution sandbox, an agent can create files, delete directories, and run commands without directly affecting real systems. The sandbox accumulates the work; a human explicitly approves what is exported. Isolation does not replace the need to check the output before it escapes. section 5.6.6

Different boundaries solve different problems:

  • Isolated cloud environments have their own filesystem, runtime, and network boundaries. DevContainers use containers; microVMs such as Firecracker provide hardware-enforced boundaries. section 5.6.6
  • Ephemeral environments give each pipeline run a fresh clone, separating one run’s work from the next. section 5.6.6
  • A Conda environment isolates installed libraries from other projects. This is dependency isolation, not a substitute for an execution boundary around untrusted commands. section 4.2 section 5.6.6

A refactoring workspace

For AI-assisted modernization, copy the original codebase into a separate repository and remove the copied Git metadata. Unlike a branch in the original repository, the new repository avoids inheriting its policies and workflows. Review operational assets such as CI/CD workflows and hooks for later replication. A separate repository supports experimentation, but should not be assumed to contain arbitrary commands. section 2.3.1 section 5.6.6

Listing 2.1 Copy the repo recursively .
cp -r /path/to/source/kitchensink /path/to/destination/ #A
cd /path/to/destination/ #B
rm -r .git/ #C

The listing shows the recursive copy and removal of the copied Git metadata. Check which other legacy files the copy includes. section 2.3.1

Treat sandboxing as one part of an operational safety system. Agent workflows also need observability, rollback procedures, incident response, and quality gates; isolation alone does not determine whether exported work is good. section 5.7 section 5.6.6

How Vibe Engineering covers it

How deep it goes
Explains isolation options and human-approved export, while stopping short of a detailed execution-sandbox setup.
Examples
Uses a separate modernization repository and a Conda environment to illustrate distinct forms of isolation.
What you can do after
Operate/configure a Conda environment and prepare a separate repository for AI-assisted refactoring.
How it connects
Connects isolation with observability, rollback, incident response, and quality gates in agent workflows.

Also explained in

Compare how 10 books explain sandboxing →

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Vibe Engineering. The section links point to the parts of the book it draws on.