Run Your Day from the Dashboard: How to Monitor AI Workflows in Real Time

Summary

The steps

1. 1

   Open the Dashboard and read the four summary cards

   The Dashboard is the home screen of DAVE. Navigate to it from the main navigation. You will see four summary cards arranged across the top of the page, each showing a single count beneath its label. These four numbers are your operational snapshot for the entire tenant.

   The four cards are:

   • Active Agents (blue): the count of agents whose status is not retired or disabled. This tells you how many AI agents are available to be called by running workflows. If this number drops unexpectedly, a workflow that depends on a missing agent will fail when it reaches that node.
   • Workflows (purple): the total number of workflow definitions in your tenant. This is a configuration count, not an execution count. It tells you how many distinct workflow graphs exist, not how many are running.
   • Running Instances (green): the count of workflow instances currently in the running status, actively executing nodes. This is the number to watch throughout the day. A higher number than expected may mean a backlog is building; a lower number than expected may mean instances are failing before they start or are stalling in a waiting state.
   • Pending Tasks (amber): the count of tasks created but not yet completed, typically because a Human Review node or User Interaction node is waiting for a person to act. This is your human bottleneck indicator.

   When the tenant is new or no data exists, cards display a dash instead of a number. If you see dashes on an active tenant, refresh the page: the Dashboard loads data on navigation.

   Do this

   • Log in to DAVE and navigate to the Dashboard.
   • Read all four card values before doing anything else. Note the Running Instances count and the Pending Tasks count.
   • If any card shows a dash on an active tenant, refresh the page to reload the counts.
   • If Running Instances is zero and you expect workflows to be running, check the instance list for instances stuck in waiting_user_input, waiting_review, or paused status. These are not counted as running.

   Example

   You log in on a Tuesday morning and see: Active Agents 8, Workflows 12, Running Instances 3, Pending Tasks 11. The ratio of 11 pending tasks to 3 running instances is the signal. Eleven tasks waiting for human action against only three active runs means your review queue has built up, probably from runs that completed overnight and left tasks unresolved. Your first action is the Pending Tasks panel, not the Running Instances panel.

   Best practice

   • Check the Dashboard first thing when you log in, before opening any other page. The four cards tell you whether the day is normal before you invest time in anything else.
   • Compare Running Instances to Pending Tasks every time you check. If many instances are running and few tasks are pending, workflows are progressing smoothly. If Pending Tasks is high relative to Running Instances, human review is the bottleneck.
   • Do not rely on memory for the counts. The Dashboard does not auto-refresh: navigate away and back, or refresh the page, to get current numbers.
2. 2

   Use the Recent Instances panel to spot problems in live runs

   The Recent Instances panel sits in the lower-left of the Dashboard. It lists the most recently created or updated workflow instances in your tenant. Each entry shows the instance's status badge, colour-coded to match the status reference: green for running, amber for paused or waiting_user_input, pink for waiting_review, red for failed, blue for completed, gray for draft or cancelled.

   The panel is a quick-scan surface. You are not reading it in detail: you are looking for red badges (failed instances) and amber or pink badges that have been sitting for longer than expected. A waiting_user_input or waiting_review instance will not proceed until a person acts on the pending task it generated. If those badges are accumulating, the workflow is stalling at a human step.

   Click any instance row to open the instance detail page, where you can see the full task table, the context viewer, and the control buttons for that specific run.

   Do this

   • Scan the Recent Instances panel for red badges. A failed instance needs investigation: open it, expand the Context viewer, and check the tasks table for the node where the failure occurred.
   • Scan for amber or pink badges. If an instance has been in waiting_user_input or waiting_review for longer than your expected SLA, open it and find the pending task in the tasks table. The instance will not move until that task is resolved.
   • If an instance is in paused status and you did not pause it intentionally, open the detail page and click Resume to restart execution, or Cancel if the run is no longer needed. Cancellation is permanent and cannot be undone.
   • If the panel is empty and you expect instances to exist, the tenant may be new, or all recent instances may have completed and scrolled out of the recent window. Navigate to the full instance list for a complete view.

   Example

   The Recent Instances panel shows five entries. Three are blue (completed), one is green (running), and one is red (failed). You click the red entry. The instance detail page opens. The tasks table shows one task with a failed status on a node labelled agent_interaction. You expand the Context viewer and find an error message indicating the AI provider returned an authentication error. The fix is to navigate to Admin, open the Providers page, and update the API key for the affected provider. Once updated, you can create a new instance from the same workflow version to rerun the job.

   Best practice

   • Treat a red badge in the Recent Instances panel as a priority interrupt. Failed instances rarely fix themselves: they need a human to diagnose and either rerun or escalate.
   • Use the Context viewer on the instance detail page as your first debugging tool. It shows the accumulated data at the point of failure, which is usually enough to identify whether the problem is a provider issue, a data issue, or a workflow design issue.
   • Prefer Pause over Cancel when you are unsure. A paused instance can be resumed; a cancelled instance cannot be restarted from where it stopped.
3. 3

   Clear the Pending Tasks panel before it becomes a backlog

   The Pending Tasks panel sits in the lower-right of the Dashboard. It lists tasks that have been created by running or recently completed workflow instances but have not yet been resolved. These are the human action items generated by your AI workflows: review requests, approval decisions, and user input prompts that the workflow is waiting on before it can continue.

   A task in this panel means a workflow instance is paused at a human decision point. The instance will not proceed until the task is completed. If tasks accumulate here without being cleared, your running instance count will eventually drop as workflows stall, and your Pending Tasks count will rise, which is the pattern that signals a human review bottleneck.

   The panel is empty when no tasks are pending. The empty state message reads: "No tasks pending. Tasks will appear here when workflow instances require review or input."

   Do this

   • Open the Pending Tasks panel each time you check the Dashboard. If it contains entries, work through them in order of age: oldest tasks first, since those represent the longest-stalled workflow instances.
   • Click a task entry to open the task detail page and complete the required action: submit a review decision, provide user input, or approve the pending item.
   • After clearing a task, return to the Dashboard and refresh the page to confirm the Pending Tasks count has decreased and the associated instance has returned to running status.
   • If the Pending Tasks panel is consistently full at the start of each day, the volume of review tasks is outpacing your reviewer capacity. Consider whether Human Review node criteria can be tightened to reduce unnecessary review triggers, or whether additional reviewers need access to the workspace.

   Example

   The Pending Tasks panel shows seven entries. You click the oldest one. The task detail page opens and shows a Human Review task generated by a content-moderation workflow. You review the flagged output, mark it as approved, and submit. You return to the Dashboard, refresh the page, and see Pending Tasks drop to six and Running Instances increase by one as the workflow instance that was waiting on that review resumes execution.

   Best practice

   • Clear pending tasks daily, ideally at the start of the working day, so workflow instances do not stall overnight into a second day.
   • If a task has been pending for longer than your expected SLA and the assigned reviewer is unavailable, an Admin or a user with the Use role can complete tasks. Admin, Create, Curate, and Use roles can all click through to the Tasks destination. Reporting and Financial roles cannot.
   • Use the Pending Tasks count on the Dashboard as a leading indicator, not a lagging one. A rising count over several days means a process problem, not just a busy day.
4. 4

   Know which roles can see and act on what

   The Dashboard is visible to all six default roles in DAVE: Admin, Create, Curate, Use, Reporting, and Financial. Every user who logs in sees the four summary cards. However, clicking a card navigates to the corresponding section, and those destination pages enforce their own role restrictions.

   The role access rules for Dashboard card destinations are:

   • Active Agents card: navigates to the Agents page. Accessible to Admin, Create, and Curate roles.
   • Workflows card: navigates to the Workflows page. Accessible to Admin, Create, and Curate roles.
   • Running Instances card: navigates to the Instances page. Accessible to Admin, Create, Curate, Use, and Reporting roles.
   • Pending Tasks card: navigates to the Tasks page. Accessible to Admin, Create, Curate, and Use roles.

   A user who clicks a card their role cannot access will receive a 403 Forbidden response. The card itself is always visible: the restriction applies to the destination, not the card.

   For daily operations monitoring, the most important access consideration is the Pending Tasks panel. A user with the Reporting role can see the Pending Tasks count on the Dashboard and click through to the Running Instances destination, but cannot click through to the Tasks destination to resolve pending tasks. If you have team members whose job is to clear the review queue, they need at minimum the Use role, not just Reporting.

   Do this

   • Confirm that everyone who monitors workflows daily has at least the Use role. Reporting-only users can see counts but cannot resolve pending tasks or control instances.
   • If a team member reports a 403 error when clicking a Dashboard card, check their role assignment. The card destination, not the Dashboard itself, is enforcing the restriction.
   • Roles in DAVE are editable. If the six default roles do not map cleanly to your team's responsibilities, an Admin can adjust role permissions. The six defaults are: Admin, Create, Curate, Use, Reporting, and Financial.

   Example

   A new team member is assigned the Reporting role so they can check workflow health each morning. They log in, see the Dashboard, and notice Pending Tasks shows 9. They click the Pending Tasks card and receive a 403 error. The fix is to update their role to Use, which grants access to the Tasks destination and the ability to resolve pending tasks, without granting access to workflow configuration or agent settings.

   Best practice

   • Assign the Use role to anyone whose daily job includes resolving pending tasks or starting workflow instances. Reporting is a read-only role for metrics, not an operations role.
   • Do not assign Admin to team members who only need to monitor and clear tasks. Admin grants access to provider configuration, credential management, and tenant settings. Use is sufficient for daily workflow operations.
   • Review role assignments when team responsibilities change. A user who moves from a monitoring function to an operations function needs their role updated before they can act on pending tasks.

What does the DAVE Dashboard tell you about your AI workflows?

How do you read the four Dashboard cards together?

Which instance statuses need your attention during the day?

What does a good daily Dashboard monitoring rhythm look like?