Agent Reliability Grader
8 questions on how you run agents in production, then a letter-graded scorecard with the one thing to fix next. Every dimension links to how.
Runs in your browser. Copy your grade to share it.
How do you run agents?
Do you score agent outputs for quality, not just check that the run completed?
Do you keep a fixed golden dataset of known-good inputs and outputs?
Do you re-run evals on every prompt and model-version change?
Do you watch for quality drift over time, with no code change?
Do you know the cost per agent, per task, not just a total bill?
Do you get alerted when failure or cost rates spike?
Do high-risk runs route to a human before they act?
Can you block or hold an agent action in real time when it breaches a threshold?
The scorecard
Answer the questions above to see your grade and where to start.
The dimensions you scored low are the ones Prefactor covers in one place: scoring, drift, per-agent cost, and holding a bad action before it lands. Dev tier: 25,000 free spans a month.
How the grade is made
Eight equal dimensions. Yes scores full, partly half, no zero; the percentage maps to a letter, and your lowest dimension is the fix to make next. A self-assessment, so only as honest as your answers.
These eight are the loop Prefactor closes in one place. The platform →
Frequently asked questions
How is the grade calculated?
Eight equal dimensions. Yes is full, partly is half, no is zero; the percentage maps to a letter, and your lowest dimension is the fix to make next. All client-side.
What is a good score?
Most teams land C to B: they score outputs and watch cost but skip regression checks, drift, or runtime enforcement. An A means the full loop is closed.
Why does each dimension link out?
So a weak score is a reading list, not just a letter. Each links a guide or glossary entry on the practice.