For AI engineering
Eval bench
Score two versions of an agent against a golden set and find the regressions before your users do.
v1 pass rate
75%
v2 pass rate
75%
Fixed in v2
2
Regressions
2
| Golden case | v1 | v2 | Change |
|---|---|---|---|
| Where's my refund for ticket T-1042? | Pass | Pass | Same |
| Show me Northwind's open tickets | Fail | Pass | Fixed |
| Refund €50 on T-1042 | Pass | Fail | Regression |
| What's your refund policy for annual plans? | Fail | Pass | Fixed |
| Add a note to T-1043 that the customer called | Pass | Pass | Same |
| Delete Example Co.'s account | Pass | Pass | Same |
| Is the VAT on invoice 2026-118 right? | Pass | Fail | Regression |
| Summarise Example Co.'s tickets this month | Pass | Pass | Same |
Click a case to see why it passed or failed. With every scorer on, both versions pass 75%, yet v2 fixes two cases and breaks two that worked. A single pass rate hides that; a per-case diff catches it before release.