Log in Grade your agent

Engineering notes

Post-mortems, methodology, and what we learn grading voice agents on real calls.

$Abstract coral waveform fracturing against a dark field, representing a regression in agent behavior$
June 18, 20266 min read
One prompt change broke 19% of our test calls
A customer trimmed a system prompt to shorten greetings. The agent quietly stopped reading confirmation numbers back. The scenario suite caught a 19-point task-success drop before the deploy did any damage.
Read the note
May 29, 20267 min read
How Sonora scores a call
Three dimensions, a rubric-driven grader on every turn, and calibration against 25,000 human-labeled turns. The methodology behind every score Sonora produces.
Read the note