Check what the reply says
A semantic check asks whether the reply explains the approval policy.
AI agent testing from the traces you already emit. See what happened in your tracing tool, then see whether the execution broke a rule in Mibo.
Traces give saved tests the evidence to check your rules. In the example below, the refund went through and the manager approval check failed.
| Observability | Testing with Mibo |
|---|---|
| See which steps ran and which tools the agent called. | Check whether the recorded actions followed your rules. |
| Inspect a trace for the sequence of events and the values recorded. | See the verdict and the evidence behind a check. |
| Read the trace to investigate what happened. | Pick a recorded trace. AI drafts a test for you to review and save, then Mibo runs it on later eligible traces. |
| Collect and inspect traces. | Use those traces as evidence for tests. |
A semantic check asks whether the reply explains the approval policy.
A procedural check uses the recorded refund amount and approval step to check the rule.
Point your agent’s OTLP/HTTP exporter at the OpenTelemetry Collector and configure both destinations. Enable message content capture in your instrumentation so spans include usable input or output. Run your agent and confirm the trace appears in Mibo with the evidence you need before drafting a test.
See both setup optionsreceivers:
otlp:
protocols:
http:
endpoint: 127.0.0.1:4318
exporters:
otlp_http/mibo:
traces_endpoint: https://api.mibo-ai.com/public/traces
headers:
x-api-key: ${env:MIBO_API_KEY}
otlp_http/other:
traces_endpoint: '<YOUR_OTHER_OTLP_HTTP_TRACES_ENDPOINT>'
service:
pipelines:
traces:
receivers: [otlp]
exporters: [otlp_http/mibo, otlp_http/other]Keep cost and latency dashboards in your observability stack and prompt management in your development workflow. Your app runs the agent and enforces its rules. Mibo evaluates traces after execution, outside the request path.
If your existing evaluations already check the tool-call rules you care about on production traces, you can keep that setup.
See how a test drafted from one recorded execution checks later traces within plan limits.
Explore AI agent testingSee how semantic and procedural checks use recorded evidence.
Explore evaluation