Traces · Cost
Two views. Same source of truth.
One agent run becomes a hierarchical trace and a cost record. Both queryable, both replayable, both exportable.
{ "files_changed": 6, "lines_added": 184 }Agent spend, last 7 days
$342.18 · ↓ 12% vs. prior week
Pillars
The eight signals every agent team needs.
Pick the one you're missing. We probably ship it.
Distributed traces
Every tool call, every LLM hop, every retry — nested as OTel-compatible spans you can replay.
Real-time metrics
Latency, throughput, success rate, token spend — per agent, per team, per tenant, per minute.
Reasoning visibility
See the chain-of-thought that led to a tool pick. Diff it across runs. Catch model regressions early.
Failure forensics
Auto-classified failure modes — wrong tool, hallucination, timeout, guardrail trip — with rerun playgrounds.
Cost analysis
Per-tenant + per-team budgets. Forecast overruns before they happen. Throttle on overshoot.
Replay & diff
Replay any trace with a different model. See what changes. Promote winners to production.
Audit & compliance
Immutable run history. Per-tenant retention policy. SOC2-friendly access controls.
Continuous learning
Feed validated runs back into memory + prompt-tuning loops. Improve without retraining.
OpenTelemetry-native
Ship spans where you already look.
OTLP exporter is built-in. Your spans land in Datadog, Honeycomb, Grafana Tempo, or your own collector — no extra agent, no proprietary format.
- OTLP gRPC + HTTP, both ends.
- Standard OTel semantic conventions, plus agent + model + tool attributes.
- Per-tenant resource attributes so your warehouse partitions cleanly.
- Zero-config sampling that prioritises failed + slow runs.
Trace production beta agent runs.
Join the waitlist for full traces, metrics, approvals, and cost telemetry during the open beta.