product / 04 · observability
See every step an agent took, and what each one cost.
Each run produces a full trace: a tree of spans for model calls, tool calls, guardrail checks, and approvals, with token counts, cost, and latency on every node. Filter your fleet by outcome, cost, or policy hits. Replay any run step by step. Export everything over OpenTelemetry to the tools you already use.
reading the figure
What you are looking at
-
Each bar is one span. Length shows time, so the slow step is obvious at a glance.
-
Every model span carries its tokens and its price, so cost is never a monthly surprise.
-
The bottom line totals the run: 4.21 seconds, 12,916 tokens, and less than half a cent.
how it works
Three steps, one trace.
-
step 01
Trace
Every run records a tree of spans with tokens, cost, latency, and outcome on each node.
-
step 02
Roll up
Tag runs with your own metadata and get cost and outcome totals per tenant, feature, or experiment.
-
step 03
Export
Send the same spans to your existing dashboards over OpenTelemetry. Nothing is locked in.
capabilities
What observability gives you
04.1
Span-level cost accounting
Input tokens, output tokens, cache reads, and dollars on every model span, rolled up per run, per agent, per tenant, and per feature using your own metadata.
04.2
Step-through replay
Rebuild any run: what the model saw, what it said, what the tool returned, and what the guardrail decided. A 3 a.m. incident review takes minutes, not a log dig.
04.3
Fleet views and alerting
Outcome rates, p95 latency, and cost per resolved run, sliced by agent version. Get an alert when error rates spike or cost per run drifts past a threshold you set.
04.4
OpenTelemetry out
Traces export as standard OTel spans to Datadog, Grafana, Honeycomb, or your own collector. Trellis is a source of truth, not a data prison.
when teams reach for it
Where it pays off first
- Teams who cannot answer "why did the agent do that?" today
- Finance asking what each customer's AI usage costs
- On-call engineers who need a replay, not a pile of logs
- Platforms that already run a tracing or metrics stack
What it costs
sample pricingPick a depth, pay a flat adder
Sampled traces (10% kept for 14 days) are free. Full traces add $0.0004 per run for 100% kept 30 days. Forensic adds $0.0009 per run for 90 days with full payloads.
questions
Observability, answered plainly
Can I send traces to Datadog or Grafana?
Yes. Every span exports over OpenTelemetry with the standard GenAI attributes, and the integrations gallery has ready-made dashboards for both.
How long do you keep traces?
It depends on the tier you pick per workspace: 14 days sampled, 30 days full, or 90 days forensic. Changing retention is logged in the audit trail.
Can I see cost per customer or per feature?
Yes. Tag runs with keys like tenant or feature, and Trellis rolls up cost, latency, and outcomes per key. Enterprise invoices can split line items the same way.
$ trellis estimate --honest
Try observability on a real workflow.
Price the stack in the configurator, then bring one of your own workflows to a demo. We'll trace it live and show you this module on your data.