Agent Mode
Agent Mode transforms Side Chat from a conversational assistant into an autonomous AI partner. Unlike Ask Mode, the Agent independently explores your codebase, creates or modifies files, and executes terminal commands—all while keeping you informed and in control.
Advantages of the Agent Mode
Solving Container Resource Issues: Because the agent runs without needing heavy local installations, it completely resolves the resource drain and performance bottlenecks typically caused by running resource-intensive CLI tools within the container.
Perfect Synergy with Arkain's Ecosystem: The AI Agent seamlessly integrates with the existing environments and powerful features provided by Arkain. By fully leveraging these native capabilities, the agent delivers far superior and more accurate results for your projects.
Access to Existing Side Chat Features: You can still utilize all the useful features provided by the existing Side Chat while using the Agent Mode, making your workflow even more convenient and fluid.
How the Agent Works
Once you provide a requirement, the Agent follows a transparent, step-by-step workflow:
Code Exploration: The Agent searches your files using pattern matching and content search to understand the project structure.
Action Plan: The Agent generates a plan displayed as a checklist above the input area, allowing you to track its progress in real-time.
Thinking Process: Between actions, the Agent's reasoning is displayed in real time, so you can follow its approach and decision-making.
File Changes & Diff View: When the Agent modifies or creates files, the changes appear instantly in the Diff Viewer. You can inspect every addition or removal before they are finalized.
Terminal Execution: If a terminal command is required, the Agent presents the command for your review before execution.
Web Fetching: The Agent can access external URLs to retrieve live documentation, API specifications, or technical updates to ensure the implementation follows the latest standards.
Control & Safety: Reviewing Actions
The Agent requests your approval before executing terminal commands:
Accept / Reject: Each command suggestion shows [Accept] and [Reject] buttons. Click [Accept] to allow execution, or [Reject] to decline.
Command Results: Once a command is executed, the terminal output is displayed directly within the conversation for your review.
Stopping the Agent: Click the [Stop] button at any time to immediately halt the Agent's current operation.
The conditions under which the approval dialog appears can be customized through Agent Permissions below.
Agent Permissions
You can configure fine-grained permissions for the Agent through the "Agent Permissions" section in Preferences. All permission rules can be managed via the GUI. Your permission settings are saved and persist across conversations.
1. Approve All
When Approve All is enabled, all actions are performed without asking. Deny rules always take precedence — even with Approve All enabled, denied commands are always blocked.
2. Tool Permissions
Control the Agent's access to file, search, and execution tools. Each tool can be set to Allow (default) or Deny. When a tool is set to Deny, the Agent is blocked from invoking it — any attempt to call the tool will be intercepted and rejected before execution.
Explore: Keyword search, file pattern matching, and directory listing
Read: File reading and code analysis
Write File: File creation
Edit File: File editing
Command: Terminal commands (including destructive operations)
Web Fetch: Fetch and analyze external web content
3. Web Domain Restrictions
Users can manage access to external web content by defining allowed and denied domains in Preferences.
Allowed Web Domains: URLs matching these domains (e.g.,
*.github.com) are accessed without confirmation.Denied Web Domains: URLs matching these domains (e.g.,
142.250.0.*) are strictly blocked.Any attempt to access a strictly blocked domain is intercepted and rejected immediately to ensure secure browsing.
4. Command Permissions
Command permissions use a cmd:args pattern format (e.g., git:, rm:, npm:install) and follow a strict priority order:
Allow patterns: Commands matching an Allow pattern are immediately executed without an approval dialog.
Deny patterns: Commands matching a Deny pattern are immediately rejected
Unmatched commands: If a command does not match any pattern, an approval dialog will appear. Click [Accept] to allow execution, or [Reject] to decline.
Note on UI Interaction
Input fields for "Allow" patterns or domain lists are automatically disabled if Approve All is active or if the corresponding tool is set to Deny, as these configurations become redundant when the tool is globally permitted or strictly blocked.
5. File Access Restrictions
You can define file path Deny patterns using glob syntax (e.g., /.env, /secrets/**).
When the Agent attempts to access a file that matches a Deny pattern, the operation is blocked and the Agent receives an error, prompting it to find an alternative approach.
This protects sensitive files such as environment variables, credentials, and secret configurations from unintended access.
Customization
Under Preferences → AI Agent → Customization, you can tailor the agent's behavior and environment to match your project's specific requirements.
1. Custom Instructions via ARKAIN.md
You can provide specific context and tailored instructions to your agent simply by adding an ARKAIN.md file to your project. Guide your agent to perform tasks exactly the way your project requires!
How to Edit: Click the [Open ARKAIN.md] button in the Custom Instructions section to open and edit the file directly within the editor.
💡 ARKAIN.md Example
Create an ARKAIN.md file in your project's root directory. The Agent will read it automatically and follow the instructions when generating code.
2. Suggested Questions
Get smart suggestions for your next question based on the agent's response. This helps you keep the conversation flowing seamlessly and efficiently.
How to Configuration: Use the toggle switch next to Suggested Questions to easily turn these automated recommendations on or off.
Available Modes: Suggested Questions are provided in Agent and Ask modes. In Plan mode, suggested questions are not provided by default.
Last updated