Skip to main content
After messages are ingested, Moda runs a set of analyses over each trace — the whole agent run, end to end, not any single API call — and across your whole workspace. This page describes each analysis at the product level: what it produces and where to find it. For the underlying objects, see the data model.

Where each result appears

Trace summaries

Each trace with at least two messages gets a one-paragraph summary. Summaries are regenerated as the trace grows, feed the semantic search index, and appear on the trace detail page and in the recent-activity section of GET /overview.

Use-case clusters

Moda splits each trace into segments — contiguous stretches of one user goal — and groups segments into a hierarchical taxonomy of use cases: broad user areas at the top, specific user goals underneath, each with a label, description, keywords, and trace counts.
  • The first taxonomy for a workspace is built once 50 or more analyzed segments exist; before that, the Clusters page shows an empty state.
  • New segments are assigned to the existing taxonomy as they are analyzed, so cluster volumes stay current.
  • The hierarchy itself is rebuilt automatically as your traffic evolves — you never configure or trigger it. Segments that match no existing use case are held aside and grouped into new nodes at the next rebuild, so a genuinely new use case becomes visible once that rebuild has run.
Use cases appear at Observe → Clusters, in the “What the agent handles” section of the Home page, and through GET /clusters, GET /clusters/:nodeId/conversations, and GET /task-clusters/search, which resolves a phrase like “refund requests” to a cluster node_id by name or meaning (CLI: moda clusters --search).

Emotion signals

Moda analyzes user messages for expressed emotion using a fixed taxonomy of 7 families and 17 signals: Five families are negative: frustration, sadness, confusion, anxiety and trust. positive is positive. disapproval is neither: it marks a calm rejection of something the agent produced, with no annoyance in it (“no, not that one”), never counts toward the negative rate, and carries no risk score. Each detection records a score, a risk score (negative families only), a trajectory (none, building, peak, resolved, sustained), an elicitor — whether the agent (bot), an external factor, the user themselves (self), or something ambiguous caused the emotion — a primary cause, and verbatim user quotes as evidence. The Signals → User Emotion tab reports the negative rate (per user turn by default, or per task or trace) and affected users. Emotion analysis runs once a trace has been quiet for about 30 minutes, and at least hourly on a trace that keeps going; traces escalated for deeper negative-emotion analysis can take longer to finish. Read them at Signals → User Emotion or via GET /emotions.

Tool-failure classification

Every tool call is recorded as it is ingested — name, success or failure, error text, and an effect classification (read, write, send, spawn, compute) — which powers the Tool Catalog within seconds. Failed calls are then classified into failure categories. Every workspace draws on the same set of categories, and Signals → Tool Failures lists the ones your failed calls fall into. Each category carries:
  • a label — a short human-readable name for the failure
  • a failure family — the broader class the failure belongs to
  • a likely cause — what most plausibly produced the error
  • an action hint — what to try first when fixing it
Failed calls are classified once the trace has been quiet for about an hour, and at least every four hours on a trace that keeps going. Classified failures appear at Signals → Tool Failures (with per-tool drill-downs and time series) and via GET /tool-failures and GET /tool-failures/:toolName.

Laziness signals

Laziness detection flags agent effort shortfalls: replies that omit code they promised (”… rest of code …”), stubbed implementations, deflection to the user, promised actions with no tool call following, claimed inability to do things the tool catalog says the agent can do, and similar patterns. Each detection carries a pattern family and a confidence:
  • Detections at or above 0.5 confidence count as laziness on their own.
  • Weaker detections count only when two or more pattern families co-occur in the same trace.
  • Truncated-data patterns are treated as ingestion artifacts, not agent behavior, and are excluded from headline counts.
Laziness detection runs once a trace has been quiet for about 30 minutes, and at least every two hours on a trace that keeps going. The Signals page has no Laziness tab: laziness reaches you through alerts (the built-in Laziness signal), laziness rates and KPI tiles in GET /trends, and the CLI’s moda traces findings and moda detections list --family=laziness.

Hallucination detection

A hallucination is an agent message that told someone something the conversation shows is false: a reply, or a message the agent sent through a tool, such as a text, an email or a LinkedIn message. Hallucinations come from Problems: when the incident validator keeps an incident as a false statement (the agent’s own words, shown false by a tool result, the user or other context in the conversation), each agent message it quotes becomes a hallucination, with the quoted claim, the evidence that shows it false, and the incident it comes from (incident_id). Hallucinations are counted, not given as a rate. Earlier detections, from a check that scored agent claims against the trace’s tool results and task status, stay as history: their contradictions count as hallucinations, and their verified and unverified claims are kept as they were. Hallucinations are written when Problems reads the trace, about an hour after its last message. They appear with the problem their incident belongs to, can drive alerts through the built-in Hallucinations signal, and are available via GET /hallucinations. The Signals page has no Hallucinations tab.

Problems

Problems group negative signals across traces by root cause, so one underlying defect shows up as one ranked item instead of hundreds of scattered detections.
  • Attribution — as each trace is analyzed, its negative signals are attributed to an existing problem or held in the unexplained remainder.
  • Discovery — problem definitions are created and updated from the attributed evidence, and the list is re-ranked as evidence accumulates.
  • Verification — attributions are audited and stamped with a verification status, and each problem maintains a dossier: executive summary, impact, evidence, and investigation reports.
  • Lifecycle — problems move through Open, Reopened, Fixed, and Resolved.
  • Feedback loop — you can mark a problem fixed, dismiss it, flag a wrong attribution, or rename it. Feedback is incorporated automatically; marking a problem fixed arms automatic reopening if the same root cause recurs.
Problems live at Observe → Problems and in the Data API problems endpoints, including POST /problems/:id/feedback.

World state

For each trace, Moda keeps a world state: the tasks the agent worked on and the facts established along the way.
  • Tasks — one per thing the user asked for, with a status (open, resolved, cancelled, idle) and a timeline of progress; a subtask names the task it belongs to.
  • Facts — key–value facts about the user and the work, where Moda’s task tracking records them, plus the user’s durable profile carried across traces.
World state is the current state only: there is no per-message history of changes. View it in the World State view of the trace detail page or via GET /conversations/:id/world-state; the earlier /replay and /snapshot views are retired and answer 410 Gone.

Next steps