Product

The agent trace, not just the LLM call.

Most observability tooling stops at the model boundary. Trefur extends one layer up — to the decisions agents make, the tools they call, the retries they fire, and the outcomes they produce. That is where the interesting questions live.

What changes when you turn it on.

These are the outcomes engineering teams shipping production agents see in the first week — no re-platforming, no rip-and- replace.

Debug with the full picture

Every decision, every tool call, every retry — captured the moment it happens. Drill from a failing customer interaction to the exact step that went wrong.

See cost the day it drifts

Tokens and dollars attributed to the step that emitted them. Catch a 40% bill spike before it lands on the CFO's desk.

Find regressions before customers do

Alert envelopes on retry rate, error rate, p95 latency, cost-per-run. Routed to your existing oncall tooling with the trace already attached.

Swap model vendors without rebuilding observability

One canonical trace shape across OpenAI, Anthropic, Bedrock, Azure OpenAI, Gemini. Vendor decisions become data.

Adopt alongside your current stack

OpenTelemetry-native. Add Trefur as a second exporter alongside Datadog, Honeycomb, Grafana. Your existing pipeline stays in place.

A connected trace, end to end.

A single root span per agent run. Every LLM call, tool invocation, sub-agent spawn, retry, and outcome is a child span. Schema-validated, OpenTelemetry-canonical, replayable.

Drill from a failing customer interaction down to the specific decision that went wrong, in the specific deployment, on the specific build.

trace · refund-bot · 4.2s · $0.018

$ agent.run [root] 4200ms

$ llm classify_intent tokens 240 310ms

$ tool lookup_order(#A-42) 200 OK 180ms

$ llm decide_refund_action tokens 410 420ms

$ tool process_refund(#A-42) 409 conflict 190ms

$ retry tool process_refund 200 OK 720ms

$ llm draft_response tokens 312 380ms

$ return 'Your refund is on its way.'

Open standards. Every framework. Your pipeline.

Trefur reads OpenTelemetry GenAI Semantic Conventions natively. Add Trefur as a second OTel exporter alongside whatever you already run, connect your AI platforms with OAuth, or turn on Claude Code's own telemetry.

Model providers: OpenAI, Anthropic, Bedrock, Gemini, Groq, MiniMax

Standard OTLP: HTTP and gRPC at /v1/traces

Fan-out friendly: keep your existing backend, add Trefur as a second exporter

What engineering teams actually do with Trefur.

Debug failing agents in production

Search by trace_id, agent name, customer interaction, or error class. Drill into the specific decision that went wrong and the prompt context the model saw.

Catch regressions before customers do

Alert on p95 latency, retry-rate spikes, cost-per-run drift, and error-rate envelopes. Route alerts to Slack, PagerDuty, Teams, or any webhook.

Investigate cost spikes

Per-trace, per-agent, per-deployment cost roll-ups. See exactly which step pushed the token count and whether it was the LLM or the tool output.

Track cost and latency across deployments

Roll up cost-per-run, p95 latency, retry rate, and error rate by build SHA and environment. See how a new deploy moved the numbers against its baseline.