Observability
Measured, not estimated
Most teams cannot answer three basic questions about their LLM traffic: what did this request really cost, what did the safety layer catch, and what did the agent actually do across a whole run. Tracing platforms answer parts of this after the fact; in/guard/out answers all three during the request, because it sits where the answers are.
Observation vs enforcement, side by side
| Question | Tracing & eval platforms | in/guard/out |
|---|---|---|
| When it acts | After the fact – traces analyzed later | In line – findings exist before the response ships |
| What a violation triggers | An alert on a dashboard | A flag, repair, or block, recorded with the request |
| Cost accounting | Token totals per call | Guard-model sub-calls priced in; the “guardrail tax” measured per stage |
| Latency attribution | Span durations | Per-stage verdict + duration, compute split from network wait, even for concurrent checks |
| Agent runs | Traces where you added instrumentation | The run graph reconstructed from the wire – model calls, tools, checkpoints, burn-down |
| Coverage | Apps with the SDK wired in | Anything that crosses the pipeline |
| Do you need both? | Yes – deep offline analysis and regression suites are their home turf | Yes – production findings make excellent new eval cases; the two compose |
What the wire can measure
Every request records a per-stage timeline – each guardrail’s verdict and duration, split into offline compute versus network wait, even when checks run concurrently. Cost is honest: the primary model call and every guard-model sub-call (judges, shields, critics) are priced into the same request and run totals, so the “guardrail tax” is a measured number, not a promise on a pricing page.
For agents, the Runs view reconstructs the whole trace: steps, tool calls with risk tiers, violations, repeated-call highlighting, and a cost burn-down – across the LLM boundary, workflow checkpoints, and the MCP gateway. Response headers carry the essentials back to your app on every call.
What you can see
Per-stage timeline
Every guardrail’s verdict, duration, and network share per request – including which stages ran concurrently. The overhead is shown, not asserted.
True cost accounting
Token pricing for the upstream call plus every guard sub-call, aggregated per request and per run. Budgets read these same numbers.
Violations & grounding reports
What was caught, by which check, with the evidence – including the answer marked up against its source for grounding findings.
The run graph
Agent runs reconstructed step by step, with tool calls, tiers, violations, and cost burn-down – one trace across all three enforcement surfaces.
Response headers
X-Guardrails-Violations, cost, and run totals ride back on every response, so your app can react without polling a dashboard.
Frequently asked questions
What is the “guardrail tax”?
The measured overhead the checks add to a request – shown per stage, per request, in the dashboard. It depends on which checks you enable and your traffic, which is why the product measures it instead of quoting one number.
Does cost include the guard models?
Yes. Every judge, shield, and critic call is priced and attributed to the request and run that triggered it. Budgets and headers use the same totals – there is no hidden spend.
Can my app react to violations programmatically?
Yes – every response carries X-Guardrails headers with the violations, cost, and run totals; blocked requests return a 422 with the reason and details in the body.
What does the Runs view show for an agent?
The reconstructed run: every step in order, the tool calls each step made with their risk tiers, violations as badges, repeated-call highlighting for loops, and the cost burn-down across the run.
Related guardrails
See the true cost and violations on your traffic
Point in/guard/out at your production traffic and watch the dashboard show what each guardrail caught, how long it took, and exactly what it cost. We are running a limited demo - sign up and we will get you in as soon as we can.