Coverage map
HIGH create_invoice 87% resolved
MED run_payroll 41% · high variance
NONE dispute_charge 8% · falls back
NONE export_to_excel 4% · inconsistent
Trace tools tell you how your agent ran. ClearViews reads your existing conversation logs and shows you what users wanted, where the agent silently fails, and what broke after your last prompt change — with zero new instrumentation.
Old analytics mapped every action to a click or an event. With agents, the user just says what they want in natural language — and the structured event stream disappears. You can see the mechanics of every run, but not whether users got what they came for.
The taxonomy builds itself from your agent's own prompt and tools — no labeling, no annotation team, no ground truth.
From Langfuse, LangSmith, or any export. PII redacted before analysis.
Intents derived from your data. Fully unsupervised.
Classify intent, cluster responses, diff behavior across prompt versions.
Intent coverage, silent regressions, unknown demand, cost per intent.
Intent + response variance form a structured, per-session view of what users wanted and how the agent handled it — the primitive every other view reads from.
Intent distribution, coverage gaps, and the unknown-intent catalog — what users want that the agent can't do.
Response variance, regression reports, and prompt-comparison diffs after every change.
Malicious and anomalous intent feed, compliance audit trail, cost per intent.
// sits above traces, below dashboards
Start on the logs you already produce. Nothing to integrate, nothing to wait for.
// taxonomy ← your data
No labeling pipeline, no annotation team, no ground-truth data required.
A worked example on a QuickBooks-style assistant — illustrative, not customer data. One pass over your logs produces the full behavioral record: what users wanted, how well the agent served them, and where it silently fails.
run_payroll: 6× the cost — routing candidate once fixed.
12 intents shift >15% (the review threshold). Most affected: create_invoice, expense. Stable: check_balance, run_payroll. → Review the affected intents before rolling the new model out.
The same behavioral record answers a different question for every team — no extra setup, just a new lens on the same data.
Rank real demand and unserved intents to drive the roadmap with evidence, not guesses.
Diff behavior across prompt and model versions; get alerted the moment resolution drops.
Identify low-variance, high-grounding intents safe to route to a cheaper model.
Flag prompt-injection, jailbreak, and anomalous patterns with confidence scores.
Intent → classification → action → outcome, with flags on regulated-topic responses.
Intents tied to abandonment and retries become early churn and roadmap signals.
You ship a prompt change and don't know what quietly broke. Catch regressions before users feel them.
You can't see what users actually want. Get intent distribution, coverage gaps, and voice-of-customer from real sessions.
You run many agents. Compare behavior, attribute cost per intent, and govern across teams from one record.
No. Sessions are redacted of personal data before any analysis, and you can start with a small sample, your staging/eval data, or synthesized sessions.
No. ClearViews runs on exports you already produce (Langfuse, LangSmith, custom). Zero new instrumentation to get your first report.
Never. Data is used only to produce your report, isolated per partner, and deleted on request.
They show how a run executed. ClearViews shows what the user wanted and whether they got it — across your whole session corpus. It complements your trace tooling.
Send a redacted sample of your logs and we'll send back a behavioral + regression report in a few days. You keep the report — no strings.