AI Chat — User Guide

Audience: All Dagen users who want to interact with AI agents to build pipelines, explore data, and manage infrastructure through natural conversation.

Overview

The AI Chat page (/chat) is the primary interface for interacting with Dagen's AI agents.

Agent behaviour in this UI is shaped by Agent Intelligence & Skills—instructions, skills (including read_skill), rules, and lessons configured at /agent-intelligence. For AI review of GitHub PRs (separate from chat), see Git Reviews. You can ask questions in plain English, attach files for context, and watch agents execute multi-step data engineering tasks on your behalf.

Layout (what you see on /chat)

The chat experience is split into a main conversation column and a control column (on smaller screens these may stack).

Main column (conversation)

1. Prompt suggestions — One-click starters to kick off common tasks; choosing one sends the text as your message.
2. Chat history — Scrollable thread; shows loading while past messages are fetched. While the agent works, the UI reflects waiting for response. New messages auto-scroll into view.
3. Chat input — Compose and send; disabled while a response is streaming so you do not interleave partial runs.

Control column (right)

1. Agent selector — Pick from available agents (built-ins + custom). Shows loading or error if the agent list fails to load. The selected agent is highlighted.
2. Model selector — Pick the LLM backing the agent from models your admin configured (Model Settings). Disabled during streaming.
3. Task history — Past sessions for the selected agent only:
   • Regenerate title — When a session is active, ask for a new title based on the conversation.
   • New chat — Clear thread and start fresh (disabled if no agent or while processing).
   • Session list — Each row shows title (or fallback like “Session [id]”) and last updated time; click to load that session’s transcript.

If no agent is selected, history explains that you must pick an agent first.

Streaming: what can appear in the transcript

Chat uses server-sent events (streaming). Besides final answers, you may see:

You can stop a long run from the UI when available—closing the stream avoids leaving orphaned client state.

Starting a Conversation

1. Open AI Chat from the navigation menu (or press / to navigate to /chat).
2. Click Select Agent in the toolbar to choose which agent to talk to.
   • Super Agent is the default — it delegates tasks to specialist sub-agents automatically.
   • Other options include the DBT Agent, Spark Developer, Metadata Discovery, and any custom agents you have created.
3. Click Select Model to pick the AI model. Each model has a provider label (e.g., "Anthropic", "Google").
4. Type your message in the input box at the bottom (placeholder: "Type something...") and press Enter or click the send arrow.

Execution Modes

The execution mode slider sets how much the agent proceeds without asking. The value is sent to the backend as execution_mode: user_guided, semi_autonomous, or autonomous. Semi-autonomous is the typical default.

Drag the slider in the toolbar to switch modes at any time during a conversation.

Safety: Destructive SQL (e.g. DROP, DELETE) is evaluated by guards; the stream may emit a consent step before execution. Prefer Guided or Semi on unfamiliar schemas or dangerous operations. For mature automation, Auto works best together with Agent Intelligence rules and lessons.

Adding Context

Click the + button next to the message input (or the paperclip icon) to attach context:

File Attachments

You can drag and drop files onto the chat or use the paperclip icon. Supported formats:

• Maximum file size: 20 MB per file
• Maximum attachments: 10 files per message

Session Management

Your conversations are saved automatically.

• New Chat: Click the New button (plus icon) in the sidebar header to start a fresh session.
• History: Click Show History in the sidebar to browse previous conversations, listed under Recent Tasks.
• Regenerate Title: Hover over a session in the sidebar and click the refresh icon to regenerate its title.
• Resume: Click any session in the history list to continue where you left off.

If no agent is selected, the sidebar shows "Select an agent to see history."

Agent Canvas

The Canvas is a side panel where agents display rich artifacts during a conversation. Click the canvas toggle button in the toolbar to show or hide it.

Artifact types that may appear:

• Use the Layout button to switch between stack, grid, and focus views.
• Use the Clear all button to dismiss all artifacts.

When the canvas is empty, it shows "AI visualizations will appear here" with a hint: "Ask the agent to create pipelines, analyze data, or explore schemas."

Destructive Operation Consent

When an agent is about to execute a potentially dangerous SQL statement (DROP TABLE, DELETE, ALTER, etc.), a consent dialog appears.

The dialog shows:

• Operation type in the title (e.g., "Destructive Operation — DROP TABLE")
• A warning message explaining what will happen
• Affected Objects — the tables or schemas impacted
• SQL Statement — the exact SQL that will be executed
• Severity level and Database type

You have two choices:

Consent flows align with Guided and Semi modes; Auto still respects destructive-operation guards where configured.

Tips for Effective Prompting

• Be specific: "Create a dbt model that joins orders and customers on customer_id" works better than "make a model."
• Mention your database: Attach a database connection as context so the agent knows which schema to target.
• Use follow-ups: The agent retains full session context, so you can say "now add a filter for active users" without repeating yourself.
• Attach files: Upload a CSV to have the agent analyze it, or attach a SQL file for the agent to review and fix.
• Start with Guided mode for unfamiliar operations, then switch to Semi or Auto once you are comfortable.

Troubleshooting