For startups →
Three lines of code and a free tier. Ship fast without flying blind.
Use cases
Different agents. The same blind spot.
Whatever your agent does, it fails the same way: quietly, with a 200 and a confident answer. Here is what that looks like in the six places teams bring Netra in.
Customer support agents
Catch the wrong answer before the customer does.
Support agents fail politely. They return a 200, sound confident, and quote a refund policy that expired last quarter. The ticket that follows is the first signal anyone gets.
- Every conversation, as one trajectoryMulti-turn sessions stay grouped, so you read the conversation rather than reassembling log lines.
- Graded on what matters hereFaithfulness to your policy docs, correct escalation, on-policy tone, no leaked PII.
- Failure modes, not ticket countsSessions cluster by intent and outcome, so the retrieval gap shows up as a pattern.
Voice agents
Trace what the caller actually heard.
A voice agent is three systems in a trench coat: speech to text, the model, text to speech. When a call goes wrong, the interesting part is usually the seam between them.
- One turn, end to endTranscription, reasoning, tool calls and synthesis land in a single span tree.
- Latency where it hurtsPer-stage timing, so you can see which leg is making the caller wait.
- Scored on the transcriptThe same evaluators you use for text, applied to what was said.
RAG and knowledge assistants
See which chunks came back — and whether the answer used them.
Most RAG failures are not model failures. Retrieval returns nothing useful and the agent answers anyway, fluently, from whatever it remembers.
- Retrieval as a first-class spanQuery, chunks, scores and the embedding call, captured on every request.
- Faithfulness, measuredGrade the answer against the context it was given, not against a vibe.
- Empty-context detectionSurface the sessions where nothing came back and the agent kept going.
Multi-agent workflows
Follow the handoff, not just the final answer.
Once a planner calls a researcher that calls three tools, the failure is almost never in the last step. It is two hops back, in a decision nobody logged.
- Sub-agent handoffs, in orderWho called whom, with what, and what came back — as one trajectory.
- Tool loops caught earlyRetry storms and circular calls show up as shape, not as a cost surprise.
- Decision-level scoringEvaluate tool choice and routing, not only the text at the end.
Document and media agents
Image, document and video pipelines, instrumented the same way.
Agents that read invoices, generate creative or process video have the same quiet-failure problem as text agents, and far fewer tools built for them.
- One instrumentation modelThe same SDK, spans and evaluators whatever the modality.
- Cost per assetAttribute spend to the document, the render or the route that caused it.
- Human-in-the-loop scoringSend the ambiguous cases to a reviewer and keep the verdict in the same dataset.
Security and compliance
Red team before release, then keep watching.
Natural-language input and real tools make the boundary between instruction and data a convention rather than a wall. A retrieved document can carry an instruction.
- Adversary emulation, open sourceAgent OPFOR probes the agent and its MCP servers for injection, tool abuse and exfiltration.
- PII detection and maskingCatch leakage in traces, and keep sensitive fields out of storage.
- Residency and access controlSeparate US and EU regions, RBAC, and an audit trail on evaluator runs.
Or start from where you are
For enterprise →
SOC 2 Type II, HIPAA and GDPR, with separate US and EU regions.
Start today
Turn 10‑hour investigations into 10‑minute fixes
Free to start. No credit card.