Prompt Regression Monitoring Agent

A prompt in production can quietly get worse when the model is updated or the kind of inputs changes, and nobody notices until users complain. On a schedule this agent runs the production prompt against its test set and a sample of recent real inputs, and compares quality and format with the last good baseline. When the pass rate drops beyond your threshold, it looks at what changed since the last good run: model version, input mix or prompt edits. It also checks the test data itself, because a bad test case can look like a regression. It drafts an alert with the failing cases and the likely cause. If it cannot tell the cause, it says so instead of guessing. You approve any rollback or prompt change. Edge case: a drop caused by bad test data leads to fixing the data, not the prompt.