Skip to main content

AI agent evaluation with evidence for each verdict.

AI agent testing from the traces you already emit. See whether the reply met your criteria and whether the agent followed your rules for tool calls.

Evaluate real executions

See why evaluated checks passed or failed using the captured reply, tool calls and arguments. Mibo also shows why a check was skipped, for example when tool evidence is missing or a required check failed.

Draft evaluations from a trace

Pick a recorded execution. AI drafts checks from the evidence it contains. Review the draft and save it to evaluate later traces when the tests apply and plan limits allow.

Keep your trace exporter

Send the same spans to Mibo and your tracing backend through an OpenTelemetry Collector. Mibo evaluates received traces outside the request path.

Two ways to evaluate an execution

Semantic checks

AI scores the response against a written criterion, such as whether it explains the approval policy.

Procedural checks

Deterministic checks inspect recorded actions and arguments, such as which tool was called and for which account. Results depend on the fields captured in the trace.

A refund without approval

Illustrative example

Recorded execution
The response says manager approval is needed. The trace also records a refund tool call for $640 without approval.
Checks
The response must explain approval; the refund must not be issued without it.
Semantic check
PASSED
Procedural check
FAILED
Learn to create a test from your trace

Read the results by stage

A stage pass rate answers one question: how often did that stage pass among its evaluated results?

Synthetic stage report Selected sample · 20 evaluated results per stage
StagePassed / evaluatedResult counts
Routing 18 / 20 · 90% 18 passed · 2 failed
Arguments 15 / 20 · 75% 15 passed · 5 failed
Tool 17 / 20 · 85% 17 passed · 3 failed
Response 16 / 20 · 80% 16 passed · 4 failed

Each denominator includes only passed and failed outcomes for that stage. Skipped, missing, and not-applicable stages are excluded; a stage with no evaluated outcomes has no rate.

Learn how to interpret evaluation metrics

Keep exploring