https://blog.openadapt.ai/Recent content on OpenAdapt BlogHugo -- 0.164.0en-usMon, 27 Jul 2026 00:00:00 +0000https://blog.openadapt.ai/posts/the-write-audit/Mon, 27 Jul 2026 00:00:00 +0000https://blog.openadapt.ai/posts/the-write-audit/Most automation teams report one success number, and that number quietly includes the runs that wrote the wrong thing and said it went fine. Here is a vendor-neutral protocol for separating the two: a four-outcome scorecard, an oracle-strength ladder, seven persistence faults worth injecting, and the sampling math to turn a week of shadow runs into a defensible bound.https://blog.openadapt.ai/posts/contribute-for-credits/Tue, 21 Jul 2026 00:00:00 +0000https://blog.openadapt.ai/posts/contribute-for-credits/An early-access program: when you hit a new way an automation silently fails, share a sanitized, de-identified signature of it and earn run credits, while every OpenAdapt user gets an engine that now refuses that failure. Raw recordings never leave your machine, and you approve every byte.https://blog.openadapt.ai/posts/2026-07-20-compile-once-govern-every-repair/Mon, 20 Jul 2026 00:00:00 +0000https://blog.openadapt.ai/posts/2026-07-20-compile-once-govern-every-repair/Reasoning through a known workflow on every run is wasteful and unsafe. Our new technical paper shows a different design: compile one demonstration into a deterministic program, repair targets when the interface drifts, and verify effects against the system of record instead of the screen. In a fault study measured end to end, screen-only checking silently accepted 75.0% of the wrong effects that actually occurred; one out-of-band record oracle cut that to 12.5%.https://blog.openadapt.ai/posts/silent-wrong-action/Fri, 17 Jul 2026 00:00:00 +0000https://blog.openadapt.ai/posts/silent-wrong-action/Screen-only verification silently passed 5 of 7 transactional fault classes — a green banner over a wrong database. We found the failure class in our own engine first, fixed five silent wrong-write modes, and built effect verification against the system of record. Measured end to end: one out-of-band record oracle cuts undetected wrong effects from 75.0% to 12.5%.https://blog.openadapt.ai/posts/the-500th-run/Wed, 08 Jul 2026 00:00:00 +0000https://blog.openadapt.ai/posts/the-500th-run/Same task, same success check, measured on openadapt-flow 0.1.0 on 2026-07-08: 100/100 for the compiled script, 20/20 for the agent. The difference was 4.9 s vs 37.5 s per run, and $0 vs $0.27.https://blog.openadapt.ai/posts/openemr-benchmark/Wed, 08 Jul 2026 00:00:00 +0000https://blog.openadapt.ai/posts/openemr-benchmark/Compiled workflows vs. a frontier computer-use agent on real OpenEMR, measured on openadapt-flow 0.1.0 on 2026-07-08: 20/20 vs 10/10 task success, 1.8x faster, $0 vs $0.55 per run in model spend — and with agent fallback, $0.029 vs $0.238 per successful run. Deterministic compilation wins on cost and latency, and never silently writes the wrong thing.