Codegen agents operate within secure, isolated sandbox environments where they can safely execute code, install dependencies, and run commands without affecting your local machine or production systems.

​
Capabilities

Each sandbox provides a controlled environment with:

  • File System Access: Read, write, and modify files within the sandbox’s temporary file system.
  • Terminal Access: Execute shell commands (bash, sh, etc.) to run scripts, linters, formatters, build tools, and other necessary commands.
  • Process Execution: Run code in various languages (Python, Node.js, etc., depending on the sandbox image).
  • Networking: Controlled network access for tasks like installing packages or fetching data (can be restricted).

​
How Agents Use Sandboxes

Agents utilize sandboxes for a variety of tasks:

  • Testing Code: Running unit tests, integration tests, or linters against the code they’ve written or modified.
  • Verifying Changes: Executing the code to ensure it runs correctly before committing.
  • Installing Dependencies: Using package managers (pip, npm, yarn) to install necessary libraries.
  • Running Tools: Executing build scripts, code formatters, or other development utilities.

​
Configuration

Sandboxes are typically configured per-agent run or defined within your Codegen settings. You often don’t need to manage them directly, but advanced configurations might allow specifying Docker images or environment variables.

Sandboxes are ephemeral; their state is not typically persisted between agent runs unless specifically configured for caching or stateful operations.