Active regression checks
Run selected scenarios against your agent before release and evaluate the response plus observable execution evidence with the tests you configure.
Run selected scenarios deliberately before release, then evaluate available production traces after real interactions with the same configured tests. Mibo makes the evidence visible without claiming complete coverage.
An agent may pass a selected scenario while a different real interaction exposes a changed route, tool call, or response. Mibo connects deliberate checks before release with trace evidence after real interactions, so teams can investigate a change without treating one result as proof of quality.
Run selected scenarios against your agent before release and evaluate the response plus observable execution evidence with the tests you configure.
After a real interaction, send an available canonical trace to Mibo. Applicable active tests are evaluated asynchronously without Mibo calling your agent again.
Interpret results through the checks you define, the instrumentation your trace includes, input classification, and the breadth of scenarios you exercise or observe.
Use active tests to deliberately exercise known scenarios, then use passive trace evaluation to see how those same checks behave against interactions your system actually handled.
01
Describe scenarios and add semantic or procedural checks for the response, tool calls, arguments, status, schema, timing, or other evidence your trace can carry.
02
Run selected inputs against the agent before shipping a change. Review the result and its evidence as a focused regression check, not as a guarantee over every possible input.
03
Send available production traces after interactions. Mibo classifies applicable scenarios and evaluates configured tests asynchronously so the team can investigate differences over time.
AI-powered checks compare a response with criteria you describe, such as being complete, grounded, safe, or aligned. The score is evidence for investigation and depends on the response and context available to the evaluator.
Deterministic checks inspect recorded facts such as node or tool calls, arguments, HTTP status, schema, token usage, and response time. Missing fields are reported as missing instrumentation rather than silently passing or failing.
Mibo supports deliberate active runs before release and passive evaluation of available traces after real interactions. Using the same configured tests across both modes helps you compare selected scenarios with observed behavior, while keeping the limits of each signal visible.
No. Your system handles the real interaction, then sends a canonical trace. Mibo evaluates applicable active tests against that trace asynchronously and does not call the agent again for passive evaluation.
Mibo uses the input available in the trace to classify which scenario-conditional tests are relevant. Universal checks can apply to every trace, while a test may be unavailable when the trace has no usable match or lacks the required evidence.
No. Results are bounded by the scenarios and checks you configure, the interactions that occur, the classification signal, and the instrumentation and fields you send. Mibo does not claim complete coverage, guaranteed quality, or automatic remediation.
Mibo reports missing instrumentation for supported checks that cannot find the required trace attribute instead of silently treating the check as passed or failed. Add the needed fields if that evidence matters to your workflow.
Trace payloads are encrypted at rest, and product access is scoped to the account that owns them. Send synthetic or appropriately governed data that fits your own privacy requirements.
Review how real traces reach Mibo, how applicable tests run asynchronously, and how missing instrumentation is reported.
Read passive testing docs →See which span fields and attributes Mibo can use for semantic and procedural evaluation.
Read the trace reference →Understand the semantic and procedural checks behind active runs and passive traces.
Explore evaluation →Investigate behavior regressions in production traces without putting Mibo in the customer request path.
Explore monitoring →Define scenarios and run deliberate checks before release with the same tests used for production evidence.
Explore AI agent testing →Run the canonical live reliability check with a synthetic customer request.
Run a live check →Follow a practical workflow for active tests, passive traces, and interpretation limits.
Read the guide →