agent-device v0.20 introduces a replay system that cuts repeated mobile QA flows from minutes to seconds by eliminating model calls for known routes. A 3-minute 18-second Claude-driven Android task replayed in 8.8 seconds with zero model calls. The post covers experiments with the --settle flag (reducing model turns by 25% and cost variance significantly), replay script mechanics with identity-binding checks, and a 36-trial repair study. Key finding: 22 of 36 repair trials passed but only 10 preserved identity evidence intact. A hybrid approach—replaying the known route then handing off to the agent—reduced median task time by 42% and cost by 46%. The author concludes that re-recording a flow can be cheaper than repairing it, and that repair quality should be measured by evidence preservation, not just pass/fail status.
