A passing run distils its path into step intents — never coordinates — offered to later runs as advice, while the agent still trusts the live screen.
A memoryless agent rediscovers the same wizard on every run, which is robust and wasteful; a recorded script is fast and brittle. Route memory is the deliberate middle: a passing run is distilled into what it was trying to do at each step, never into coordinates, and later runs treat it as advice while still being told to trust the current screen.
Successful paths are distilled into step intents.
Coordinates are never stored — they shift between renders.
Later runs receive prior paths as advisory guidance.
The agent is told to trust the live screen over the notes.
A redesigned page degrades gracefully instead of breaking.
No. It stores what each step was trying to achieve, not where it clicked, and the agent may ignore it.
The notes stop matching and the agent falls back on the live screen, so the run degrades gracefully.
The agent spends fewer turns rediscovering a path it has already completed once.
A scenario is an objective, a role and assertions — no selectors, nothing to rewrite after a redesign.
Creator, approver and supplier each get a fresh, independently authenticated session.
A lost agent is never filed as a product defect.
Every model call is metered and aggregated from call to suite.
See the complete product: problem, features, how it works and deployment.
Request a demo to watch a multi-role regression suite run through your real interface.