Observe
Understand agent behavior through traces, spans, and sessions.
Risicare's observability layer provides deep visibility into your AI agent's behavior.
Overview
Traces
End-to-end request traces
Spans
Individual operations
Sessions
User interaction sessions
Agents
Multi-agent observability
Data Model
Session (user interaction)
└── Trace (single request)
└── Span (individual operation)
├── LLM Call
├── Tool Execution
├── Agent Decision
└── Child Spans...
Key Metrics
| Metric | Description |
|---|---|
| Latency | P50, P95, P99 response times |
| Token Usage | Prompt and completion tokens |
| Cost | USD cost per trace/session |
| Error Rate | Failed traces percentage |
Dashboard Views
Trace List
View all traces. The list has a search box, a Status control and an Errors control. The first page lists traces from all time; Load More and export use the time range of the top bar. The list has no filter for agent, session, latency or cost — see Filtering and Search.
Trace Detail
Deep dive into a single trace:
- A span waterfall, a timeline and the raw JSON
- Span attributes and events (prompt and completion text appears among the attributes when content capture is on)
- Tool inputs and outputs
- Error details and stack traces
- Total cost of the trace
Analytics
Aggregate views:
- Trace volume over time
- Error rate trends
- Latency trend
- Request counts by model
- Cost by agent
- Errors by taxonomy category
Freshness
A trace usually appears in the dashboard within seconds of the call. The dashboard does not push new traces to an open page: reload the trace list to see them. The KPI strips refresh every 30 seconds.
Filtering and Search
There is no query language
Risicare has no field:value search syntax and no boolean operators. A query
like status:error AND agent:planner is treated as one literal string and
will match nothing.
Search
The search box takes plain text and matches on exactly two things:
- the root span name, as a case-insensitive substring
- the trace ID, as a prefix
It does not search agent names, models, content, or error codes.
Filters
Filtering is done with separate controls, not search terms:
| Filter | Values |
|---|---|
| Status | Completed or Error |
| Errors | All Traces or Errors Only |
| Time window | Last hour, Last 24 hours, Last 7 days or Last 30 days, in the top bar. It applies to Load More and export, not to the first page |
These controls are on the traces page. The Sessions and Agents pages have their own time-range control, which also has an All time option and starts at Last 24 hours.
The Management API is not available with an API key during the beta, so you cannot filter traces over HTTP.
Data Retention
Retention is a fixed ClickHouse TTL that varies by data type, uniform across every project during beta:
| Data | Retention |
|---|---|
| Traces, spans, sessions | 90 days |
| Prompt/completion content | 90 days — stored on the span itself |
| Evaluations, scorer results | 365 days |
Captured prompt and completion text lives in the span's attributes, so it is retained for the span's full 90 days — not the 30 days stated by earlier versions of these docs.
The retention_days project setting is read-only: the API always reports the enforced 90, and a PATCH sending that key is rejected with 422. Per-plan retention tiers are not configurable today. See Data Management for the full table.