Agents generate code, interact with filesystems, and run shell commands. Because we can’t predict what an agent might do, it’s important that its environment is isolated so it can’t access credentials, files, or the network. Sandboxes provide this isolation by creating a boundary between the agent’s execution environment and your host system.
In deep agents, sandboxes are backends that define the environment where the agent operates. Unlike other backends (State, Filesystem, Store) which only expose file operations, sandbox backends also give the agent an execute tool for running shell commands. When you configure a sandbox backend, the agent gets:
- All standard filesystem tools (
ls, read_file, write_file, edit_file, glob, grep)
- The
execute tool for running arbitrary shell commands in the sandbox
- A secure boundary that protects your host system
Why use sandboxes?
- Security: Code runs in isolation - can’t access your credentials, files, or network
- Clean environments: Use specific dependencies or OS configurations without local setup
- Reproducibility: Consistent execution environments across teams
- Flexibility: Switch between cloud providers or use local VFS for development
How sandboxes work
Isolation boundaries
All sandbox providers protect your host system from the agent’s filesystem and shell operations. The agent cannot read your local files, access environment variables on your machine, or interfere with other processes. However, sandboxes alone do not protect against:
- Context injection — An attacker who controls part of the agent’s input can instruct it to run arbitrary commands inside the sandbox. The sandbox is isolated, but the agent has full control within it.
- Network exfiltration — Unless network access is blocked, a context-injected agent can send data out of the sandbox over HTTP or DNS. Some providers support blocking network access (e.g.,
blockNetwork: true on Modal).
See security considerations for how to handle secrets and mitigate these risks.
The execute method
Sandbox backends have a simple architecture: the only method a provider must implement is execute(), which runs a shell command and returns its output. Every other filesystem operation — read, write, edit, ls, glob, grep — is built on top of execute() by the BaseSandbox base class, which constructs scripts and runs them inside the sandbox via execute().
This design means:
- Adding a new provider is straightforward. Implement
execute() — the base class handles everything else.
- The
execute tool is conditionally available. On every model call, the harness checks whether the backend implements SandboxBackendProtocol. If not, the tool is filtered out and the agent never sees it.
When the agent calls the execute tool, it provides a command string and gets back the combined stdout/stderr, exit code, and a truncation notice if the output was too large. For example:
4
[Command succeeded with exit code 0]
bash: foobar: command not found
[Command failed with exit code 127]
read_file to access it incrementally. This prevents context window overflow.
Two planes of file access
There are two distinct ways files move in and out of a sandbox, and it’s important to understand when to use each:
Agent filesystem tools — read_file, write_file, edit_file, ls, glob, grep, and execute are the tools the LLM calls during its execution. These go through execute() inside the sandbox. The agent uses them to read code, write files, and run commands as part of its task.
File transfer APIs — the uploadFiles() and downloadFiles() methods that your application code calls. These use the provider’s native file transfer APIs (not shell commands) and are designed for moving files between your host environment and the sandbox. Use these to:
- Seed the sandbox with source code, configuration, or data before the agent runs
- Retrieve artifacts (generated code, build outputs, reports) after the agent finishes
- Pre-populate dependencies that the agent will need
Security considerations
Sandboxes isolate code execution from your host system, but they don’t protect against context injection. An attacker who controls part of the agent’s input can instruct it to read files, run commands, or exfiltrate data from within the sandbox. This makes credentials inside the sandbox especially dangerous.
Never put secrets inside a sandbox. API keys, tokens, database credentials, and other secrets injected into a sandbox (via environment variables, mounted files, or the secrets option) can be read and exfiltrated by a context-injected agent. This applies even to short-lived or scoped credentials — if an agent can access them, so can an attacker.
Handling secrets safely
If your agent needs to call authenticated APIs or access protected resources, you have two options:
-
Keep secrets in tools outside the sandbox. Define tools that run in your host environment (not inside the sandbox) and handle authentication there. The agent calls these tools by name, but never sees the credentials. This is the recommended approach.
-
Use a network proxy that injects credentials. Some sandbox providers support proxies that intercept outgoing HTTP requests from the sandbox and attach credentials (e.g.,
Authorization headers) before forwarding them. The agent never sees the secret — it just makes plain requests to a URL. This approach is not yet widely available across providers.
If you must inject secrets into a sandbox (not recommended), take these precautions:
- Enable human-in-the-loop approval for all tool calls, not just sensitive ones
- Block or restrict network access from the sandbox to limit exfiltration paths
- Use the narrowest possible credential scope and shortest possible lifetime
- Monitor sandbox network traffic for unexpected outbound requests
Even with these safeguards, this remains an unsafe workaround. A sufficiently creative enough context injection attack can bypass output filtering and HITL review. General best practices
- Review sandbox outputs before acting on them in your application
- Block sandbox network access when not needed
- Use middleware to filter or redact sensitive patterns in tool outputs
- Treat everything produced inside the sandbox as untrusted input
