AI ReliabilityPRODUCT LLM-OBSERVABILITY
LLM Observability
Traces that survive a long agent run
OpenTelemetry in, Claude semantics preserved. Accurate four-bucket cost, a transcript view that stays readable when a span tree would not, and a replay button that re-runs a captured call so you can diff the output.
What it does5 CAPABILITIES
The capability, stated plainly
- 01
Four-bucket token accounting
Input, output, cache creation and cache read are billed at four different rates. Tools that report two cannot tell you what a request cost, and they hide the cache-hit ratio — the most actionable efficiency number on an LLM workload.
- 02
A transcript, not just a span tree
Span trees are the right shape for a request and the wrong shape for a ten-minute agent run with fifteen tool calls and sub-agents. The transcript view reads in order, start to finish.
- 03
Deterministic replay
Capture enough of a run — prompt, params, model, tool inputs and outputs, retrieved context — to re-run it and diff the result against the original.
- 04
Redaction before egress
The collector runs in your infrastructure and masks fields at the edge, so sensitive payloads never reach us. It is Apache-2.0, so you can read the code that does it.
- 05
OpenTelemetry as the wire format
No proprietary SDK. The same instrumentation points at any OTLP backend, so leaving costs a configuration change rather than a migration. That is deliberate.
Why it lives here
One pipeline, not another agent to install
LLM Observability reads the same instrumented stream as everything else on the platform. That means one collector in your infrastructure, one redaction policy applied before anything leaves it, and one usage bill — rather than a separate vendor, agent and contract per question you want answered.