Looker starts running the 'which slice moved' hunt before you open the dashboard

The number went red on the dashboard on a Sunday evening, and I spent forty minutes clicking around before I worked out that a single distribution center had shipped without the promo code applied. Metric alerts had done their job: they told me a KPI moved. They just could not tell me why. That gap, the "so which slice, exactly" gap between an alert and an answer, is what Looker is trying to close with Agentic Workflows, a preview capability available in Looker version 26.08 and later.

What actually landed

Looker Agentic Workflows is a set of background agents that watch a metric and, when a threshold is crossed, automatically run a root-cause pass and drop the result into Slack or email. The root-cause pass is Key Driver Analysis, or KDA, run over the underlying data model to isolate which dimensions moved the metric. The delivered artefact is a diagnostic summary, with a direct link back to Conversational Analytics in Looker so you can keep asking follow-up questions in the same natural-language interface.

Setting up a monitor by talking to it

The setup path is where I felt the DX shift. Rather than clicking through the alerts UI to pick a Look, a field, a comparison and a delivery, you tell the agent what you want in a sentence. The examples in the announcement are "Monitor return rates weekly" and "Notify me if average order value exceeds $1,000". The agent interprets that, confirms the threshold conditions it inferred, and generates a workflow configuration plan for you to review before it starts running. That "confirm before I launch this" beat is the important part. It keeps a fuzzy natural-language request from becoming a monitor you did not realise you had.

The message that arrives in Slack

When a monitor trips, you do not just get a line about metric X moving by Y percent. You get a diagnostic summary, a short KDA writeup of which slices contributed most to the change, and an inline link that opens the same investigation in Conversational Analytics. From there you can chat the follow-ups: exclude a region, break by channel, compare against last week. In practice that turns an alert into an investigation starter instead of a Slack ping you go straight past to silence.

Who can turn it on

Two permissions gate all of this: chat_with_agent and create_alerts. If your seat does not have both, you cannot build a workflow. Business users can view and edit their own active monitors. Looker administrators have oversight across the whole instance, with the ability to review, adjust or disable workflows created by anyone. Activation itself is a per-instance toggle: an admin opens the Gemini in Looker settings page and flips on the Agentic Workflows preview toggle. Without that flag on, none of the above is visible.

Where this sits next to other patterns

The pattern here is not unique to Looker. Agent-driven RCA on top of dashboards is showing up across observability tooling, with the rough shape being: alert fires, an agent decomposes the change, the alert message contains the decomposition. What is worth noticing about the Looker take is that it happens inside the BI seat, where the person who owns the metric definition is also the one setting up the monitor, in the same conversational surface they already use to write reports. That collapses the usual handoff between the data team and the on-call engineer for the softer, business-metric monitors that never really belonged in the SRE alert pile in the first place.

The rough edges

It is preview, so the usual caveats apply. Two things I would test carefully before you build a team habit around it. First, KDA is only as good as the data model it runs against; a poorly modelled orders explore will produce a diagnostic that sounds confident and is wrong, and the natural-language wrapper does not fix that. Second, natural-language monitors are easier to create than to prune, and the sentence "Notify me if average order value exceeds $1,000" has to survive a definition change six months from now without silently going stale. Admin oversight is going to matter more than it sounds; the review-and-disable path exists for a reason.

What I am watching next

The bit I actually want to see is what happens when an agent-generated monitor and a hand-built LookML alert on the same metric disagree, and which one your team ends up trusting after a week. Agentic monitors are on a fairly obvious arc towards being the default in BI, and this preview is the first time I have felt the workflow-setup step get out of the way rather than merely add another chat pane. If you have flipped the Agentic Workflows toggle on your instance, I would love to hear which monitor was the first one you actually kept.