Diagnostics

In the Analytics group next to Insights, open Diagnostics. The diagnostics page lets you review agent and automation settings, available execution observations and possible improvements.

Viewing requires access to analytics, agents and automations. Starting a paid AI analysis also requires permission to manage analytics.

Selecting entities

In the first selector, choose what to check: Entire project, Automations or Agents.

For agents and automations, a second selector appears. It defaults to All agents or All automations: diagnostics selects a group only from that type. To check one entity, enter its name in search and click the agent or automation. To return to selection by type, choose “All agents” or “All automations” again.

Changing the type clears the specific entity selection. The selected type is retained in the report and when adding entities to the check. In “Entire project” mode, the second selector is hidden and the group can include both types. The filter limits entity selection; connected agents and nested automations are still considered when reviewing the selected scenario.

You can manually select disabled agents and draft automations as well. Executions are not required: their settings are reviewed, while execution data may be unavailable.

Choose the observation period: the last 7 or 30 days. The group is filled from the candidates found in this order:

1. New errors. Entities qualify when an error within the selected period was absent from the previous AI analysis data. If there was no previous analysis, any error within the period qualifies. The newest errors come first, including automation incidents that have already been closed.
2. Unreviewed settings. If places remain, enabled agents and running automations with a published version qualify when their current settings have not been reviewed. The most recently changed come first. Executions are not required. Settings may have been created or changed before the selected period. Editing a draft when a published version exists does not count as changing the reviewed version.
3. New activity. Remaining places go to entities with events within the period that were absent from the previous analysis data. Entities never reviewed come first, followed by those with the oldest review. Successful and failed executions count as activity. More executions do not increase priority.

Each entity takes one place. There are no separate quotas for the three groups: for example, when you select 10 entities and enough have new errors, all 10 places go to entities with errors. With two entities with new errors, three with unreviewed settings and enough candidates with new activity, the group is 2 + 3 + 5.

Only settings and observations with a completed AI analysis count as reviewed. Opening a group does not mark anything as reviewed. A partially supplied graph does not establish a full configuration review. An error or settings change during AI analysis remains new for the next selection. Disabled agents and drafts without activity are not selected automatically. If history is unavailable or incomplete, the period is not marked as reviewed. Reviewed intervals are retained separately: after a seven-day review, expanding to 30 days can discover older errors. If the added period has no new observations, changing the period alone does not require another AI analysis.

In the “Check with AI” block, enter the number of entities in the numeric field below the heading, from 1 to 50. The adjacent count shows how many matching entities are available, such as “10 of 37”. It counts entities with new data to review. When the full list size is not yet known, the number has a “+” sign. Selecting a specific agent or automation reviews one entity. AI requests full settings during the check.

If no matching entities have been found yet and the search is unfinished, use Continue searching. It continues selection without AI analysis or credit charges.

Changing the type, specific entity, period or quantity automatically updates the selection. An overview without a started AI review lasts 15 minutes. Saved results for matching data are reused.

Known errors

Errors are marked on entities in the main list. Select an agent or automation to view the details in its card.

Agents show errors within the selected period. Automations show errors within the period and open incidents separately; the latter may have occurred before that period. An already reviewed error alone does not return an entity to the AI group: another selection requires a new error, a change to the reviewed settings or new activity. The error count is not a failed execution rate. Viewing this data is available without AI analysis or credit charges.

Starting AI analysis

While the check runs, it shows the current stage, a progress bar and elapsed time. The stages are preparation, reviewing settings and activity, and checking connections and drawing conclusions. Stages advance as the check proceeds; the timer shows elapsed time, not a countdown.

After entering the quantity, click Check at the bottom of the block. In the analysis dialog, choose a model and set the total spending limit. Credits already spent on this check count toward that limit. Review how many credits are needed to start, then click Start check. The total cost depends on how much needs reviewing and will stay within your limit.

The check automatically continues reviewing the selected group until complete or until the remaining budget cannot cover the next AI request. The final cost depends on the selected model’s pricing and actual usage. When the limit is insufficient, progress is saved: increase the limit and select “Continue check”. The model of a started check remains fixed.

Starting the check opens a report without selection filters. It shows the results and the number of reviewed entities. If more entities remain, finish the current analysis, enter how many to add, then select Check more. Before analysis starts, the dialog shows the actual number of new entities, model and spending limit. Earlier results remain in the report; AI reviews only the added entities. Total report spending is retained, and previously reviewed entities are not charged again. The period and entity type stay the same.

To change the scope or period, select New check. This opens the filters for a new selection. You can also return using the “Diagnostics” link in the “Diagnostics → Report” header.

If the check's data has expired, the same button appears in the message on the page or in the model selection dialog. It prepares a fresh entity selection using the current filters. AI analysis must be started separately.

If a run has stopped and cannot be resumed, the report shows “Check stopped”, the reason and a “New check” button. The progress bar is hidden because the check is no longer running. The button opens a new selection; paid analysis starts only after confirming the model and spending limit.

Select Pause to take a break. The part already running will finish and be charged. Closing the page also pauses the check. The started check, original data, findings and spending are retained for 30 days after the last stage. The URL reopens it within the same user session. Opening a saved check is free; resuming AI analysis requires clicking the button.

When settings and observations match a saved result, it is reused. The record of reviewed settings and events is stored separately: report expiry alone does not return an unchanged entity to automatic selection. If AI has responded but saving the result failed, retrying the same check resumes the saved operation without another AI request. If the last AI request’s outcome is unknown, another paid request is blocked.

When a check stops because data is unavailable, the message identifies the source when known. Restore access and select “Continue check”: previously read data, findings and spending are retained. New AI requests are charged within the selected limit. If settings changed during reading, select “New check” and start a review of the current version. When the check stops because of data volume or analysis duration, do not automatically repeat it: send its link to support. Saved findings remain available but do not mean the entire scenario was reviewed.

Reading results

Select an entity in the group. Its result card shows activity statistics, findings and suggested actions. The counters reflect recorded errors but do not explain their causes on their own. Until AI finishes the analysis, the card indicates that results are not yet available or that the review is incomplete.

Each finding initially shows its main point and suggested action. Select Details to expand the evidence, quotes and instructions for verifying the outcome. AI findings include verified exact quotes and source names, such as “Agent instruction: Support”. Quotes are saved with the result and passed to the assistant when discussing it. They show which text supports the finding; the AI conclusion still needs evaluation.

• A possible problem is a cause or risk suggested by AI. It needs verification.
• A recommendation suggests how to improve settings or the workflow.

Missing execution data does not mean the entity is working correctly. If only part of the history could be retrieved, the card says so; unknown values are not presented as zeros. Original settings are read through the existing user MCP with your access permissions. Diagnostics is available through a Cabinet user session. AI chooses and calls the permitted MCP read methods itself: it obtains the complete graph, including every step’s settings, nested automations and the full instructions of connected agents; it requests variables by ID or name. To assess operation, AI can request examples of automation executions and agent events within the selected period, then examine a specific execution in detail. This sample does not cover all history. Large responses are passed to AI in pages without shortening instruction text. After reading every page, AI separately compares the entire scenario: what data one step creates, what the next receives, and whether agent instructions, branches and loops agree. When working context fills up, notes referencing original sources are saved; originals remain available for rereading. Until reading and the final relationship review finish, the scenario remains unreviewed. If a source is unavailable or an analysis limit is reached, this is shown separately, without a conclusion about the scenario’s health. The card also shows the number of events, step runs and errors in the selected period, and the current number of open errors for an automation. No findings does not guarantee that there are no problems either.

Instruction wording is checked against the intended use of variables: {{lead.phone}} can correctly insert a value, but does not always tell an agent where to write a new phone number. An @ reference does not replace write permission either. Diagnostics compares the wording with actual bindings and permissions; it does not label every {{...}} expression as an error. Passwords and access keys are hidden. External services are not executed. Knowledge-base content and the current state of every lead and dialog are not automatically loaded; operational conclusions apply only to the examined examples.

For automations, diagnostics reviews the published version. If none exists, it reviews the draft and marks this in the card. Errors from another version are not evidence of a defect in the current workflow.

All findings are saved. The card initially shows two; select Show more findings to expand the rest.

Next steps

Click Open settings to navigate to the entity. The button is available with the corresponding permissions. Alternatively, click Discuss with assistant for an explanation of the findings and a plan of changes.

Diagnostics does not apply changes automatically. After changing settings, select “New check”, choose the agent or automation and start a review to evaluate the outcome.