Skip to main content
Two inventories are built entirely from ingested traces, with nothing to configure: Users (who is using your agent and how it is going for them) and the Tool Catalog (every tool the agent calls and how often its calls succeed).

Users

Open Observe → Users. Users are keyed by the user ID you attach to ingested events, so the page fills in once traces arrive with user IDs set (see Ingestion overview). User aggregates refresh hourly.

User list

  • KPIs: Total Users, Active 7d, At Risk, Avg Health, plus charts for active users over time and the health-score distribution.
  • Cohort pills filter the list: power users, heavy tool users, frustrated, at risk, new this week, and silent sufferers. Clicking an active cohort clears it.
  • Search matches user IDs; wrap a term in double quotes ("user_123") for an exact match.
  • Time window presets: Last 24h, 7d, 30d, 90d.
  • Table columns: User, Traces, Health, Score, Top Intent, Signal, Last Seen. Page sizes 10/25/50/100 with numbered pagination.

Health score

Each user carries a health score from 0 to 100 for the selected window, shown in the Score column and in the detail page header. It reflects how many of the user’s traces went wrong, weighted by how badly. Each trace takes a penalty, capped at 1, from what was detected in it: 0.6 for frustration, 0.3 for laziness, 0.2 for a problem found in the trace’s context, 0.15 for confusion, disapproval, anxiety, sadness or distrust, and 0.1 per failed tool call, up to four. The score is 100 × (1 − total penalty ÷ (traces + 1)), kept between 0 and 100. The extra trace in the divisor acts as one clean trace, so a single bad trace scores 50, not 0. A hallucination counts when it surfaces as a Problem; a contradicted claim on the Signals page does not lower the score on its own. The Health column is a band on that score: Healthy at 70 or above, At Risk from 40 to 69, and Frustrated below 40. This is a different number from the 0 to 1 Confidence on the Signals page, which measures how strongly the detector saw a given emotion, not how the user is doing.

Per-user history

Click a user to open their detail page (fixed 30-day window):
  • Header with identity and health status, and a stats row: Traces, Turns.
  • Callouts — notable findings for this user, such as behavior concentration and frustration trajectory.
  • Intent distribution and Frustration trajectory charts.
  • An Activity feed of the user’s traces with a load-more control; each row opens the full trace.

Tool Catalog

Open Harness → Tools. Every tool call in ingested traces produces a record within seconds, so the catalog is close to real time.

Tool inventory

  • Search by tool name, filter by effect or by outcome (has failed calls, or all calls succeeded), and pick a time range from Last 24 hours to Last 90 days.
  • Columns: Tool name, Calls, Success, Traces, Effect, System, Last seen. Sort by Calls, Success, Traces, or Last seen.

Effect classification

Each tool is classified by the kind of effect it has: Tools whose effect cannot be determined show ?.

Tool detail

Clicking a tool opens its detail page:
  • What the model reads — the tool’s registered description, when one is registered.
  • Input parameters — registered parameters merged with parameters observed in real calls, with types, call percentages, and common values.
  • Common uses — what the agent uses the tool for.

Failure stats

The Success column is each tool’s share of recorded tool receipts that succeeded. Signals → Tool Failures counts classified failures and divides by recorded receipts for its KPI. moda detections tools divides classified failures by classified events. These rates can differ from the catalog. For failure categories — what kind of error, the likely cause, and a suggested action — use Signals → Tool Failures, or the Data API’s GET /tool-failures and GET /tool-failures/:toolName.

From the CLI or the Data API

The same catalog is available over an API key: moda tools list and moda tools get <tool_name> in the CLI, or GET /tools and GET /tools/:name. Both read the same queries as this page, and GET /tools/:name also returns the top failure categories and the registered schema.

Next steps

  • Signals — tool-failure categories, emotion, laziness, and hallucination detections.
  • Traces — the trace detail page and its World State view.
  • Ingestion overview — how to attach user IDs so the Users page populates.