Monitoring

Monitor agent traffic, latency, errors, tokens, model cost, and tools over time.

Monitoring turns trace data into time-series charts. Use it to spot changes in traffic, latency, errors, token usage, model cost, and tool behavior across many agent runs.

Use Debug traces when you need to investigate one request.

Choose a time range

Monitoring supports:

Presets from the last hour through the last 30 days
All retained time
Relative ranges such as 15h or 1d
Custom dates from the calendar picker

The default range is the last 7 days.

Group the data

Group charts by:

Model
Run name
Tool
User
Tag
Release

Choose None to view the project as a whole. Grouping is useful when one project contains several agents, models, releases, or customer segments.

Read the dashboard

The selected time range and grouping apply across the dashboard.

Traces

Chart	Meaning
Trace Count	Number of traces over time.
Trace Latency	P50 and P99 trace latency.
Trace Error Rate	Percentage of traces that ended in error.

LLM calls

Chart	Meaning
LLM Count	Number of generation, model, and embedding observations.
LLM Latency	P50 and P99 latency for LLM observations.

Cost and tokens

Chart	Meaning
Total Cost	Total model cost over time.
Cost per Trace	Median and P99 cost per trace.
Input Tokens	Total input tokens.
Input Tokens per Trace	Distribution of input tokens per trace.
Output Tokens	Total output tokens.
Output Tokens per Trace	Distribution of output tokens per trace.

WeaveScope uses explicit event costs when available and otherwise calculates cost from model pricing .

A cost increase with stable trace volume often points to longer prompts, more output, a model or service-tier change, or newly emitted explicit costs.

Tools

Chart	Meaning
Run Count by Tool	Tool calls over time.
Median Latency by Tool	Median tool latency.
Error Rate by Tool	Percentage of tool calls that failed.

Run types

Chart	Meaning
Run Count by Name	Runs grouped by trace name.
Median Latency by Run Name	Median latency for each trace name.
Error Rate by Run Name	Error rate for each trace name.

Feedback scores

Feedback Scores shows average evaluation, annotation, feedback, or monitor scores when that data is attached to traces or observations.

Understand missing data

Monitoring includes only data still inside the project's retention window. Recently accepted events may also need a brief processing delay before they appear in charts.

If an individual trace is available but a chart has not updated yet, use the trace detail for immediate debugging and return to Monitoring for aggregate trends.