Independent effect check
How do I know an RPA bot or computer-use agent actually saved the record?
OpenAdapt compiles a demonstration into a program. The program reports VERIFIED only if an independent check agrees. Run openadapt quickstart --break-it to see the failure an RPA bot or computer-use agent can hide: the application paints success, the backend does not save the record, and the independent read returns HALTED.
pip install openadapt
openadapt quickstart
openadapt quickstart --break-it
The healthy quickstart writes a synthetic record through the GUI and confirms it through a separate read-only interface. The --break-it run uses a backend that lies. Same compiled program. Different oracle result. A Saved banner is not the check.
When not to use OpenAdapt for this
If an API exists, use the API, and let that API be the write and the check. A computer-use agent still fits a task you have not demonstrated. Screen-only checking is the trap: in a 90-run injected-fault study it accepted 54 of 90 wrong results. A complete system-of-record read accepted 0 of 90. Read the study
What a halt looks like
The GUI says the note was saved. The independent read does not find the row. The program stops and keeps the report. It does not retry blindly if delivery might already have happened; that outcome is RECONCILIATION_REQUIRED. Ask any bot or agent vendor for those counts, not only task-success rate.
Compare the tools that skip this check
Playwright, Power Automate, UiPath, and Claude computer use can all click Saved. None of them is a substitute for an independent read of the record.