Reset
Start each task from the same declared precondition.
Design-partner validation
Know what changed.
Prove what broke.
Runstance is validating deterministic regression evidence for browser agents: reset the task, repeat the run, grade the resulting state, and gate the release.
Not a finished product. No credentials or production access.
Regression evidencesample
TASKUpdate billing address
BLOCKv0.8.45 / 5
v0.9.03 / 5
Final state mismatch
Agent reported success, but address.postcode was not persisted in 2 of 5 runs.
For teams shipping
Browser agentsComputer useWorkflow agentsAgent QA01 / Method
Trace success is not state success.
A confident completion message can hide a wrong cart, a stale record, or an action that only worked once. Runstance is designed around the environment state that remains after the agent stops.
Repeat
Expose flakes and drift across versions and multiple runs.
Grade state
Check the observable postcondition, not the agent’s claim.
Gate
Ship, investigate, or block with evidence attached.
02 / Problem
Built for the failures demos skip.
The validation targets the gap between task completion and dependable release evidence.
Wrong final state
The trace looks plausible, but the environment says otherwise.
Version regression
A new model or prompt improves the average and breaks a critical task.
Run-to-run variance
One lucky pass hides instability that appears only under repetition.
03 / Validation pilot
Two paid commitments required to proceed
A narrow test before a product.
We are looking for teams with a real missed regression and a safe way to reproduce it. The design-partner pilot only starts after scope, payment, and data handling are agreed.
Discuss a qualifying taskDESIGN PILOTUS$500
14 days- 10 resettable tasks
- 3 versions or models
- 5 repeats per configuration
- 1 evidence report and runbook
Refundable before kickoff. No promised outcome, certification, or production integration.
04 / Boundaries
Bring a redacted trace. Keep the keys.
No production credentials
Only a redacted trace or isolated sandbox is in scope.
Thirty-day deletion
Technical artifacts are deleted within 30 days or sooner on request.
No training or publicity
No model training, public benchmark, logo, or case study without separate written permission.
One regression is enough to start