AI agent issues

The Issues board is where data problems get tracked. Two things land on the same board:

• AI-filed findings — every AI agent turn is scanned in the background; likely-wrong answers are grouped by root cause and filed automatically.
• Human-filed issues — an org admin can file an issue directly from the board.

Each issue has a priority, an assignee, an activity feed, and — where applicable — a one-click writeback that opens a pull request against your dbt project or your shared project context file.

​

When to use it

Turn Issues on once you have agents in regular use and want a shared queue for data-quality work. It’s most useful when:

• Users ask the same kind of question repeatedly and the agent struggles
• You want to keep dbt metadata and agent instructions tight as your data model evolves
• You’re rolling agents out to a wider audience and want admin-level visibility into quality
• You want a single place to triage data problems users spot in dashboards or charts

If you’re still setting your first agent up, start with agent setup and evaluations first — Issues works best once there’s real usage to learn from.

​

Enabling Issues

Issues is opt-in. Until an admin enables it, no agent turns are processed and no findings are collected.

1. Go to Settings → Ask AI → General.
2. Toggle Review AI agent turns on.

Once enabled, future agent turns are scanned in the background and human-filed issues can be reported from anywhere in the app. Existing threads are not back-filled. To stop AI scanning, toggle the setting off. Previously collected issues remain on the board until you act on or dismiss them.

​

Opening the board

Open Settings → Ask AI → Issues to see the board. The Ask AI button in the navbar is the fast path: when there are open issues it previews recent findings; when the board is empty it shows a small dropdown that links straight to the Issues board so newly-filed manual issues are always reachable. Cards on the board are grouped by status lane (To Do, In Progress, Done) and each card shows its priority, assignee, and category badge. Click a card to open the issue detail modal.

​

How issues get onto the board

​

AI-filed findings

With the setting enabled, Lightdash scans every AI agent turn for signs the answer was probably wrong — a missing metric, an ambiguous field, a gap in your project context — and files an issue for each root cause it identifies. Scanning doesn’t depend on who asked. Turns from users who don’t have the permissions to open pull requests through AI writeback or the AI coding agent are scanned too, so what went wrong in their conversations still reaches the board, where someone with those permissions can fix it. The classifier works from the stored record of the turn — it runs no queries of its own. See Privacy and data handling for what that record contains. The record includes the turn’s thumbs-up or thumbs-down rating and, for a thumbs-down, the comment left with it. Explicit feedback counts as evidence in its own right, so a thumbs-down can lead to a finding even when nothing else in the conversation signals a problem. The rating is stored on the agent’s answer without the name of the person who gave it. The person who started a thread can rate its answers, and so can anyone with the manage:AiAgent scope in the project, such as admins and developers. Before filing, the classifier checks the board’s existing issues and, if the same underlying problem is already tracked, attaches the new evidence to that issue instead of opening a duplicate. Recurring findings bump the Recurs N× tag on the existing card. Slack notifications (if configured) only fire when a new issue is opened. Recurrences accrue silently on the existing card so a chatty problem doesn’t spam the channel — check the board for the latest state. AI findings are opened in the same lane your triage workflow uses (existing findings behavior is unchanged).

​

Human-filed issues

Any organization admin with AI agents enabled can file an issue from the New issue button on the Issues board. Manual issues land in the To Do lane.

​

Filing an issue

The Create issue modal captures: Submit and the issue appears on the board in To Do, with the author recorded as the creator.

​

The issue detail modal

Clicking a card opens a focused modal with a metadata rail (status, priority, assignee, agent, project) and a chronological activity feed that merges issue events with any existing writeback events:

• created
• status_changed
• assignee_changed
• priority_changed
• comment_added
• recurred

You can change status, priority, and assignee inline; each change writes an event to the feed with its author. comment_added events appear inline in the feed too, but the modal has no comment box. To comment on an issue that has a linked pull request, open Project settings → Pull requests and click View thread on that pull request. For AI-filed findings, the modal also has an Evidence section that shows the curated excerpts the classifier actually cited — not the full agent thread. A Read full conversation action opens the source thread in a stacked modal (or in a new tab) if you need the wider context. Findings also show an affected-target chip and a Recurs N× tag next to the category badge, and — if writeback is blocked — a plain-language note explaining why (for example, that a GitHub or GitLab connection is required).

​

Fixing an issue with writeback

For each issue, Lightdash proposes the smallest change that would prevent it:

• Semantic layer fixes open a pull request against your dbt or native Lightdash YAML project (rename a field, add a description, add a metric).
• Project context fixes add a short note to a shared file that your agents read before answering future questions.

Writeback runs the same way whether the issue was AI-filed or human-filed. A manual issue is eligible for writeback when it has:

• a root cause of semantic layer or project context,
• an assigned agent, and
• a project connected to GitHub, GitLab, or Bitbucket Cloud (GitHub only for project context fixes).

When an issue is eligible, a Create PR button appears at the top of the issue. An issue still in triage shows Accept and Dismiss instead, so accept it first.

• For a semantic layer issue, Create PR starts the writeback in the background, which can take a few minutes, and opens the workspace, where you follow the pull request and its verification.
• For a project context issue, Create PR opens a preview of the change. Click Looks good, open PR to confirm.

If an issue is blocked from writeback, the issue shows the reason, such as a missing repository connection. For manual issues, the Related explores you picked when filing are passed into the writeback prompt so the agent starts from the right model instead of guessing from the description. Project context fixes update agent context in your project repository. For native projects, see AI writeback and Issues for connection and validation behavior.

​

Project context vs. semantic layer fixes

Both kinds of fix open a pull request, but they change different things and have different review costs. A rule of thumb: if a new analyst joining the team would also benefit from the change, it’s probably a semantic layer fix. If only the agent needs to know it, it’s project context.

​

Sending issues to Linear or Jira

If your team plans work in Linear or Jira, Lightdash can open an issue there for every finding. New findings create an issue automatically, existing open findings can be exported in one click, and each issue links back to the finding on the board. Status is not synced between the two. Set it up from Settings → Ask AI → General, in the Linear issues and Jira issues sections that appear once reviews are on. See Send issues to Linear and Jira for the full setup and what gets exported.

​

Privacy and data handling

• Issues runs against agent turns inside your organization only.
• Issues and findings are stored in Lightdash alongside your other agent data and respect your existing project and admin permissions.
• Disabling the toggle stops new AI scanning immediately; it does not delete previously collected issues.
• The classifier reads the stored record of the turn — the user’s prompt, the agent’s answer, the agent config snapshot at the time, and the tool calls the agent made, with their saved results truncated to a preview. It doesn’t run its own queries against your warehouse.
• It respects each agent’s data access setting. With data access off, the agent’s query tools save only row counts (and column names, for SQL queries) with no row data, so the classifier never sees result rows for that agent’s turns. It still sees the prompt, the answer, and the filter values the agent used to build its query.

​

Related

• Send issues to Linear and Jira — mirror findings into the tracker your team already uses.
• Agent context — the shared lightdash.project_context.yml file project context fixes write to.
• Agent visibility — where Issues fits into admin visibility overall.
• Evaluations — run a fixed set of prompts against your agent and grade the answers.
• AI writeback — let an agent open a dbt pull request from chat.
• Autopilot — scheduled agent that cleans up content and flags issues for review.