Add Guardrail for DeepAgent CLI — Restrict File Operations to a Specified Folder and Require User Approval Before Actions

[tanghoong]

The deepagent CLI currently has the capability to create or modify files and directories during its execution. However, there is currently no built-in guardrail to ensure that these actions are limited to a safe, user-specified directory, nor any interactive confirmation mechanism before performing such actions.

This raises potential security and safety concerns, especially when the CLI is used in sensitive environments or integrated within automated workflows.

Problem / Motivation

Unrestricted File Access:

The DeepAgent may create or modify files outside the intended project directory, which could unintentionally affect unrelated system files or repositories.

Lack of User Confirmation:

The agent can perform file operations automatically without explicit user consent, which poses a risk when generating, deleting, or overwriting files.

Security & Trust Concerns:

Developers or organizations integrating DeepAgent into production workflows need to ensure that the agent acts within a clearly defined and auditable scope.

Proposed Solution

Add two guardrail mechanisms to the DeepAgent CLI:

Directory Restriction:

Introduce a CLI parameter or configuration option (e.g., --safe-dir ) that explicitly defines the maximum allowed directory scope for file creation and modification.

Any attempt to access paths outside this directory should raise an exception or require manual approval.

User Approval Workflow:

Before executing file operations (e.g., create, modify, delete), prompt the user for approval.

Example interaction:

DeepAgent wants to create: /myproject/src/agent/handler.py  
Proceed? [y/N]

Add a --yes flag for automation or CI/CD usage where confirmation is pre-approved.

Benefits

Increases safety and transparency in agent actions.

Prevents accidental or malicious file modifications.

Builds trust for enterprise users and regulated environments.

Encourages adoption by ensuring better control and auditability.

Additional Context

This feature aligns with the growing need for secure autonomous agent operations, particularly as LangChain expands into use cases involving agentic workflows that write or manage code. Guardrails like these help ensure the system remains predictable, verifiable, and user-trusted — essential qualities for scaling production-grade AI agents.