Wrong route
Check that a request reaches the right specialist and workflow, not only that the endpoint responds.
Test what your AI agent says and does.
Check whether an agent gives the right answer and takes the right steps. Start with a recorded trace and build tests around what it contains.
A successful response can hide a wrong action
Define the expected workflow, tool arguments, and response for a request. Check them separately: an agent can choose the right route and still act on the wrong account.
Tool behavior
Verify calls, arguments, and other observable facts in the agent trace.
Incomplete answer
Evaluate whether the response is clear, safe, and aligned with the outcome your customer needs.
Review the behavior contract
Define a contract, then inspect the execution.
Define the expected outcome
State the permitted action, required arguments, and what the answer may claim. Include a forbidden path and a tool-failure path.
Check the evidence
Inspect the recorded input, response, tool calls, and results. Confirm the fields required by each check are captured before interpreting an absence.
Turn a failure into a test
Use AI to draft checks from an available trace, review and save them, then evaluate later traces. Compare a known correct execution with a deliberately broken one.
See what happened in the trace
Semantic checks
Evaluate meaning, safety, completeness, and alignment with the outcome you described instead of requiring a fixed string.
Procedural checks
Inspect execution facts such as tool calls, arguments, attributes, HTTP status, and schema.
Frequently asked questions
What does Mibo test in an AI agent?
Mibo evaluates the behavior you describe, including the response and observable execution evidence such as routes, tool calls, values, and status. Rule-based checks verify facts, while AI-powered checks score responses against plain-language criteria.
Can I test an agent before and after release?
After an interaction, send its recorded execution over OTLP. Mibo evaluates applicable configured checks without calling the agent again.
Does trace evaluation add Mibo to the customer request path?
No. Your system handles the interaction, then sends a trace to Mibo. Mibo evaluates the trace asynchronously in the background.
Keep exploring
AI agent evaluation
Check the answer and the execution together.
Explore evaluationAI agent monitoring
Investigate responses and tool calls in production.
Explore monitoringAI agent test cases
Explore examples for routing, deletion, and refunds.
Explore AI agent test casesAI agent safety testing
Inspect evidence for safety-related behavior and its limits.
Explore AI agent safety testingAI agent test coverage
Find scenarios with traffic and those still idle.
Explore test coverage