Classify every trace by intent
Each new trace is matched to its closest intent in real time, so the picture stays current as traffic grows.
Classify every trace by intent: Docs (opens in new tab)Agent Insights
One bad trace is an anecdote.
A thousand bad traces is a roadmap.
Netra clusters production traffic into the intents your users actually have, and flags when cost, latency, errors or tool use drift, with no rules to write.
Trusted by teams shipping agents in production
What is Agent Insights?
A daily read on how your agent is really behaving. Insights clusters your production traces into intents, classifies every new trace, and compares cost, latency, errors and tool use against a rolling baseline. What moved is summarised in a brief, ranked by urgency, and there is nothing to configure.
See the jobs your agent is really doing
After about 500 traces, your traffic is clustered into labelled intents and every new trace is classified. You see workflows instead of averages.
Spot new intents early
Traces that match no known intent are surfaced daily, often the first sign users want something new.
Spot new intents early: Docs (opens in new tab)Open a detailed intent view
Cost, latency, step count, error rate, output length and tool use for one intent.
Open a detailed intent view: Docs (opens in new tab)Follow each intent's volume trend
A status label from the last 30 days shows which workflows are growing and which are fading.
Follow each intent's volume trend: Docs (opens in new tab)Catch the regression an average hides
Cost, latency, step count, error rate, output length and tool use are tracked per intent, so a slowdown in one workflow can't hide behind the rest.
Detect drift automatically
Behavioural, output and input drift, scored with calibrated statistics. No thresholds to tune.
Detect drift automatically: Docs (opens in new tab)Notice shifts in tool usage
Changes in which tools an intent calls, measured with Jensen–Shannon divergence. It is the quiet change behind many regressions.
Notice shifts in tool usage: Docs (opens in new tab)Read detailed signals
Each signal names the metric, the window, the intent it affects and how far it moved.
Read detailed signals: Docs (opens in new tab)Compare against a rolling baseline
Every metric is measured against the previous 30 days, refreshed daily.
Compare against a rolling baseline: Docs (opens in new tab)Read one brief, not ten dashboards
After each observation window, a short digest of the most significant drifts and new intents, ranked so the team starts with what needs action today.
Start with a daily overview
The drifts and new intents from the latest cycle, in the order they matter.
Start with a daily overview: Docs (opens in new tab)
Sort into must act, should investigate, good to know
Every finding is ranked by urgency, so the team acts on the first and skims the last.
Sort into must act, should investigate, good to know: Docs (opens in new tab)Open a detailed insight view
The intent, the metrics that moved and the traces that show it, in one place.
Open a detailed insight view: Docs (opens in new tab)See trends across multiple timelines
Daily, weekly and monthly windows, so a slow slide shows up as clearly as a spike.
See trends across multiple timelines: Docs (opens in new tab)From raw traces to what to fix next
Take an insight into your IDE with the traces behind it, and put an alert on the patterns you never want to miss again.
Debug using MCP
Pull the traces and sessions behind an insight into Cursor, Claude Code or Windsurf through the Netra MCP server.
Debug using MCP: Docs (opens in new tab)Turn findings into test cases
Add the failing traces to a dataset, so the fix is checked on every release.
Turn findings into test cases: Docs (opens in new tab)
Set alert rules on top
Layer alert rules over insight metrics when a drift should notify someone, not just appear in the brief.
Set alert rules on top: Docs (opens in new tab)Check that the fix worked
Each new brief shows whether the drift settled and the intent's metrics moved back.
Check that the fix worked: Docs (opens in new tab)Runs on the traces you already send
Nothing to set up beyond the SDK you already use. Insights starts working as traces arrive, with defaults calibrated for every metric.
- No rules to write
- Rolling 30-day baseline
- Refreshed daily
- Calibrated defaults
- Daily · weekly · monthly windows
- Works with any Netra SDK
Start with no configuration
Baselines, windows and severity levels are calibrated for you, with no rules to write.
Docs (opens in new tab)Learn from your own traffic
A rolling 30-day baseline, refreshed daily, keeps comparisons fair as your agent changes.
Docs (opens in new tab)In production
Explain every decision your agents make
How Skydda explains every agent decision
For security workflows, explainability isn't optional. Netra let us show exactly how our agents investigate — every decision, tool call, and LLM exchange, giving CISOs the confidence they need.
Frequently Asked Questions
Everything You Need to Know About Agent Insights
Can't find the answer here? The Agent Insights docs go deeper, or talk to our team.
What does Agent Insights need from me?
Nothing beyond traces flowing through the Netra SDK. Intents, baselines and drift detection all run automatically with calibrated defaults.
How much traffic before intents appear?
About 500 traces. After that, Insights clusters your traffic into labelled intents and classifies every new trace against them.
What kinds of drift does it detect?
Behavioural drift in cost per trace, latency and step count; output drift in error rate and output length; tool-usage drift in which tools are called; and input drift when user requests stop matching known intents.
Which time windows does it compare?
Daily, weekly and monthly — each compared against a rolling baseline, with the 30-day baseline refreshed daily.
How does it decide a change is significant?
Most metrics are scored with Z-scores; error rate and output length use percentage change; tool distribution uses Jensen–Shannon divergence. Severity is calibrated for you rather than set by hand.
Can I be notified about drift?
Yes. Each observation cycle produces a brief of the most significant drifts and new intents, and you can add alert rules on top of the metrics.
Start today
Turn a thousand traces into the next thing to fix
Send traces, and let Netra tell you what your users are doing and what changed this week. Free to start.