Overview

Watch production traffic, quality, latency, cost, and the health of your improvement loop

Monitor is the live dashboard for a project. Use it to see whether your AI agent is receiving logs, how traffic is changing, where quality or runtime is drifting, and which evidence needs deeper review in Traces, Behaviors, Evaluators, Datasets, or Improve.

What Monitor shows

The top of Monitor answers the fastest operational questions:

Signal	What it tells you
Total traces	How much production traffic Adaline has received for the selected time range.
Throughput	Average request volume over time.
P95 latency	Tail latency for real user requests.
Error rate	How often requests are failing.
Average eval score	Continuous evaluation quality trend, when evaluators are attached.
Total LLM cost	Provider spend from logged model calls.

The Agent metabolism panel shows the state of the improvement loop: cycles, approved cycles, rejected cycles, in-flight cycles, average cycle time, new Behaviors, and active evals. It is not only a dashboard card. It is a quick answer to “is the agent learning from production or just accumulating logs?”

Read it as a triage surface

Start broad, then open evidence:

1. Pick the time range that matches the release, incident, or customer report.
2. Scan the KPI row for obvious movement in volume, latency, cost, errors, or eval score.
3. Review the chart groups for the metric that moved.
4. Use View traces from a chart, or open Traces, to inspect the exact requests behind the metric.
5. If the pattern repeats, check Behaviors. If the fix belongs in the prompt, start or review an Improve cycle.

Monitor should not be the final stop for important decisions. It points you to evidence, and the evidence should become a trace review, Behavior investigation, dataset row, evaluator update, or Improve cycle.

Dashboard sections

Monitor groups charts by the kind of decision they support:

Section	Use it for
Traffic & volume	Confirm logs are arriving, request volume changed, or spans per trace changed.
Performance & latency	Find slow requests, tail latency, and bottlenecks before editing prompts.
Quality	Watch eval score and pass-rate movement across production traffic.
Cost & spend	Track total cost, tokens, and whether usage changed after a prompt, model, or traffic shift.
Model breakdown	Compare cost, latency, tokens, and efficiency by model.
Environment breakdown	Separate production, staging, and other deployment environments when metadata is present.
Performance breakdown	Identify slow prompts or other high-impact surfaces.
Tool usage	See which tools or functions are being called and how often.

For chart details, see Analyze log charts.

Good Monitor data

Monitor becomes useful when your logs contain clear names, spans, status, input and output content, model usage, costs, tokens, tags, and safe metadata such as environment, route, release, customer segment, or feature flag.

If charts look empty, flat, or hard to segment, the next step is usually instrumentation, not analysis. See Integrate your AI Agent and Instrument overview.

Analyze log charts

Read traffic, quality, latency, cost, model, environment, and tool charts.

Analyze log traces

Move from a metric to the exact traces behind it.

Filter, search, export logs

Narrow production traffic, inspect matching traces, and export the result set for review.

Deep search

Find relevant logs by meaning across traces and spans.

Use logs to improve prompts

Turn production evidence into datasets, evaluators, and Improve cycles.