Prompt, context and completion
The rendered prompt, the retrieved chunks, the tool arguments and the raw completion, on every span — so a regression is read rather than reconstructed.
By agent type · Text
The answer looked fine. The reasoning wasn't.
Chat assistants, RAG systems, copilots and tool-using agents fail fluently. The response is well formed, the status code is 200, and the mistake is three spans back in a retrieval that came home empty. Netra traces the whole trajectory and grades it.
What Netra captures on every turn
Multi-turn as one trajectory
Session-level tracing keeps a conversation together. You open the session that went wrong, not fourteen disconnected requests.
Tool choice and routing
Which tool the agent picked, what it passed, what came back, and how many times it tried again before giving up.
Token, model and retry spend
Cost attributed per agent, per route and per user, so the retry storm shows up as a shape long before it shows up on the invoice.
Prompt and model lineage
Every response tied to the prompt version and model that produced it, so you can bisect a quality drop to the exact change.
Framework-agnostic instrumentation
OpenTelemetry-native, with auto-instrumentation for LangChain, LlamaIndex, CrewAI and the OpenAI and Anthropic SDKs.
The failures worth naming
Answered with no context
Retrieval returns nothing useful and the agent answers anyway, from whatever it remembers. Netra surfaces the empty-context sessions instead of burying them among the 200s.
Confidently out of date
A faithfulness evaluator grades the answer against the context it was actually given, so a fluent quote from a policy that expired last quarter scores badly.
The quiet tool loop
Circular calls and retry storms read as a pattern in the trajectory, not as a slow month-end bill.
Prove the fix before it ships
Turn the traces that broke into a regression set, grade the change against them, then rehearse it against simulated users and a red-team pass. The same evaluators keep running on live traffic afterwards, so the score you shipped on is the score you are held to.
Start today
Trace your first text agent in five minutes
Three lines of code, 5,000 spans a month free. No credit card.