What must your AI do for customers?
Describe the customer’s task, the expected result and the checks for success.
Matches use the sample library. No customer files are uploaded.
Compare versions on customer work
Test the same workflows across models and agent versions.
Cost & quality
Baseline: Zarna 1.1
Last replay results13 improved3 regressed40/50 accepted$5.10 / accepted
Zarna 1.1toZarna 1.2
- Accepted tasks
- 30/50to40/50
- Failed tasks
- 20to10
- Retries / accepted task
- 1.00to0.20
- Cost / accepted task
- $19.40to$5.10
- Total spend
- $582.00to$204.00
Simulated example replay on the same captured tasks. No live agents or customer tools are executed. Unrun agents have no chart values.
Production issues
Review workflow gaps and turn them into tests.
Example production activity · 3 of 12 review samples
Turn gaps into tests
1 test draft · ready for review
Production
Find workflow issues and turn them into tests.
Example window128 tasks 11 failed 23 retried
Last 24 h · 3 customers · 6 workflows · Zarna 1.1 · deployed
| Customer | Workflows | Tasks | Failed | Retried |
|---|---|---|---|---|
| Percheron | Diligence, Deal updates | 52 | 5 | 9 |
| Larkspur | Portfolio reporting, Cash forecasting | 41 | 4 | 8 |
| Harrow | Portfolio operations, Investment research | 35 | 2 | 6 |
34 gaps · 16 staples · 50 tasks in Production suite v1 · 10 pending de-identification
Sep 14 example window; no live source is connected. Verified includes passed and failed outcomes, not a success rate. These fixed counts do not change with sample reviews or simulated arrivals.
Workflow issues
12 to review · 12 example records