Product · Tool Validation
Did the action actually happen?
The core problem: an agent says "your refund is processed" and never calls refund_order. We check the trace, not the sentence.
- Most common failure
- Claimed, not executed
- Detection
- Trace-based
- False positives
- Near zero
Call verification
Assert that the required tool ran, in the right order, at the right turn.
Argument checking
Amount, order ID, reason code, validated against the conversation context.
Silent-failure detection
Catch tool errors the agent smoothed over in its reply.
Claim vs action diff
Flag every statement the transcript makes that the trace does not support.
Test what your agent actually does
Connect an agent, run the suite, and find the failures before your customers do.