All docs
See what's happeningIssuesdashboard

Issues

Failing checks grouped by cause or agent, ranked, each with where to correct it.

Issues turns hundreds of failing checks into a short, ranked list of causes, each with where to correct it. Use it to decide what to fix first. It covers the last 24 hours.

Use it

Issues, grouped by causeIssues, grouped by cause

  1. Choose Group by cause (one row per failing check, such as "Right version of the source") or Group by agent (one row per agent, for multi-agent apps).
  2. Read each row's stage and Where to correct. The stages are:
StageThe fix belongs inExample
RetrievalThe search or document lookupThe 2025 guide was retrieved for a 2026 question
AnswerThe prompt or model call that writes the answerA document said $1,500 and the answer said $2,500
Tool callA tool your agent called, or its argumentsThe SQL query had no plan-year filter
RoutingWhich agent or tool the request was sent toA refund question went to the billing agent
Hand-offWhat one agent passed to the nextThe onset date changed between intake and adjudication
Agent stepsThe order or number of stepsThe same lookup was called five times
SetupYour project's settingsA check needs a data definition that isn't there
Not localisedNot enough information to saySend documents, tools and agent names to narrow it

A claim that a supplied document contradicts is placed on the answer step, because the right fact was retrieved. Retrieval is named only when the documents lack the fact. 3. Act on it:

  • Open events failing … lists those answers in Live events.
  • Create alert starts an alert on that check's failure rate (Alerts).
  • Collect into dataset starts a collection rule for answers failing it, so the next release is tested on them.
  • Configure (Setup issues only) opens Data definitions.

Not configured isn't an issue

Issues lists failing checks only. A check that is Not configured is not a failure: the project hasn't given the data it needs (for example, SQL checks with no schema). Open an answer's Checks tab to see which definition it needs, then add it under Data definitions.

A worked example

Group by cause shows "SQL follows the rules · 38 answers · Tool call". Where to correct says the query had no filter on plan_year. You click Open events failing SQL follows the rules, read two queries to confirm, and add "always filter claims by plan year" to the agent's SQL prompt. Then Create alert with a threshold of 5%, so you hear about it if it comes back.