Skip to content

GitLab Duo Agentic Chat

  • Tier: Premium, Ultimate
  • Add-on: GitLab Duo Core, Pro, or Enterprise
  • Offering: GitLab.com, GitLab Self-Managed
  • Status: Experiment
  • LLMs: Anthropic Claude Sonnet 4

Version history

The availability of this feature is controlled by feature flags. For more information, see the history.

GitLab Duo Agentic Chat is an enhanced version of GitLab Duo Chat that can autonomously perform actions on your behalf to answer complex questions more comprehensively.

While Chat answers questions based on a single context, Agentic Chat searches, retrieves, and combines information from multiple sources across your GitLab projects to provide more thorough and relevant answers. Agentic Chat can also create and edit files for you.

"Agentic" means that Agentic Chat:

  • Autonomously uses a large language model to determine what information is needed.
  • Executes a sequence of operations to gather that information.
  • Formulates a response to your questions.
  • Can create and change local files.

For larger problems, like understanding a codebase or generating an implementation plan, use the software development flow of the GitLab Duo Agent Platform.

For an overview, see GitLab Duo Agentic Chat.

Use Agentic Chat

You can use Agentic Chat in the GitLab UI or in VS Code.

Prerequisites:

  • A GitLab Duo Core, Pro, or Enterprise add-on.
  • A Premium or Ultimate subscription.
  • You have an assigned seat for or access to GitLab Duo Chat.
  • You have turned on beta and experimental features for your GitLab instance or group.

Use Agentic Chat in the GitLab UI

To use Agentic Chat in the GitLab UI:

  1. In the upper-right corner, select GitLab Duo Chat. A drawer opens on the right side of your screen.
  2. Under the chat text box, turn on the Agentic mode (Beta) toggle.
  3. Enter your question in the chat text box and press Enter or select Send. It may take a few seconds for the interactive AI chat to produce an answer.
  4. Optional. Ask a follow-up question.

Use Agentic Chat in VS Code

Prerequisites:

You can only use Agentic Chat in a project:

  • Hosted on a GitLab instance.
  • That is part of a group that meets the prerequisites.

To use Agentic Chat:

  1. In VS Code, go to Settings > Settings.
  2. Search for gitlab agentic.
  3. Under Gitlab › Duo Agentic Chat: Enabled, select the Enable GitLab Duo Agentic Chat checkbox.
  4. On the left sidebar, select GitLab Duo Agentic Chat ({duo-agentic-chat}).
  5. Select Refresh page if prompted.
  6. In the message box, enter your question and press Enter or select Send.

Conversations in Agentic Chat do not expire and are stored permanently. You cannot delete these conversations.

Use Agentic Chat in JetBrains IDEs

Prerequisites:

To use Agentic Chat in a project, it must be:

  • Hosted on a GitLab instance.
  • Part of a group that meets the prerequisites.

To use Agentic Chat:

  1. In your JetBrains IDE, go to Settings > Tools > GitLab Duo.
  2. Under Features, select the Enable GitLab Duo Agentic Chat checkbox.
  3. Restart your IDE if prompted.
  4. On the left sidebar, select GitLab Duo Agentic Chat ({duo-agentic-chat}).
  5. In the message box, enter your question and press Enter or select Send.

Conversations in Agentic Chat do not expire and are stored permanently. You cannot delete these conversations.

Create custom rules

Version history
  • Custom rules added in GitLab 18.2.

If you have specific instructions that you want Agentic Chat to follow in every conversation, you can create custom rules.

Prerequisites:

Conversations that existed before you created any custom rules do not follow those rules.

  1. Create a .gitlab/duo/chat-rules.md file.

  2. Enter the custom rules into the file. For example:

    - don't put comments in the generated code
- be brief in your explanations
- always use single quotes for JavaScript strings

  3. Save the file.

  4. To have Agentic Chat follow the new custom rules, start a new conversation, or /clear the existing conversation.

    You must do this every time you change the custom rules.

Agentic Chat capabilities

Agentic Chat extends Chat capabilities with the following features:

  • Project search: Can search through your projects to find relevant issues, merge requests, and other artifacts using keyword-based search. Agentic Chat does not have semantic search capability.
  • File access: Can read and list files in your local project without you needing to manually specify file paths.
  • Create and edit files: Can create files and edit multiple files in multiple locations. This affects the local files.
  • Resource retrieval: Can automatically retrieve detailed information about issues, merge requests, and pipeline logs of your current project.
  • Multi-source analysis: Can combine information from multiple sources to provide more complete answers to complex questions. You can use Model Context Protocol to connect Agentic Chat to external data sources and tools.
  • Custom rules: Conversations can follow any customised rules that you specify.
  • Agentic Chat in the GitLab UI only - Commit creation: Can create and push commits.

Chat feature comparison

Capability Chat Agentic Chat
Ask general programming questions Yes Yes
Get answers about currently open file in the editor Yes Yes. Provide the path of the file in your question.
Provide context about specified files Yes. Use /include to add a file to the conversation. Yes. Provide the path of the file in your question.
Autonomously search project contents No Yes
Autonomously create files and change files No Yes. Ask it to change files. Note, it may overwrite changes that you have made manually and have not committed, yet.
Retrieve issues and MRs without specifying IDs No Yes. Search by other criteria. For example, an MR or issue's title or assignee.
Combine information from multiple sources No Yes
Analyze pipeline logs Yes. Requires Duo Enterprise add-on. Yes
Restart a conversation Yes. Use /reset. Yes. Use /reset.
Delete a conversation Yes. Use /clear. No
Create issues and MRs No Yes
Use Git No Yes, IDEs only
Run Shell commands No Yes, IDEs only
Run MCP tools No Yes, IDEs only

Use cases

You might find Agentic Chat particularly helpful when you:

  • Need answers that require information from multiple files or GitLab resources.
  • Want to ask questions about your codebase without having to specify exact file paths.
  • Are trying to understand the status of issues or merge requests across a project.
  • Want to have files created or edited for you.

Example prompts

Agentic Chat works best with natural language questions. Here are some examples:

  • Read the project structure and explain it to me, or Explain the project.
  • Find the API endpoints that handle user authentication in this codebase.
  • Please explain the authorization flow for <application name>.
  • How do I add a GraphQL mutation in this repository?
  • Show me how error handling is implemented across our application.
  • Component <component name> has methods for <x> and <y>. Could you split it up into two components?
  • Could you add in-line documentation for all Java files in <directory>?
  • Do merge request <MR URL> and merge request <MR URL> fully address this issue <issue URL>?

Troubleshooting

When working with Agentic Chat, you might encounter the following issues.

Network connectivity problems

Because Agentic Chat requires network access to retrieve resources, network restrictions might impact its functionality.

To help resolve Agentic Chat networking issues, see the GitLab Duo Agent Platform network issue troubleshooting documentation.

Slow response times

Agentic Chat has significant latency when processing requests.

This issue occurs because Agentic Chat makes multiple API calls to gather information, so responses often take much longer compared to Chat.

Limited permissions

Agentic Chat can only access resources that your GitLab user has permission to access, which is the same as Chat.

Search limitations

Agentic Chat uses keyword-based search instead of semantic search. This means that Agentic Chat might miss relevant content that does not contain the exact keywords used in the search.

Feedback

Because this is an experimental feature, your feedback is valuable in helping us improve it. Share your experiences, suggestions, or issues in issue 542198.

Related topics