Agent repeatedly substitutes a related-but-wrong action for the one explicitly ordered, and ignores repeated corrections

Summary

When given an explicit instruction to perform action X, the agent repeatedly performs a closely-related action Y instead — and keeps doing Y even after the user corrects it many times over several minutes. The agent acknowledges the mistake each time ("you're right, doing X now") but then either presents another substitute artifact or narrates what it is about to do, without ever producing the literally-requested observable outcome.

What happened

• I instructed the agent to run/drive a GUI application in its visible, live form and show me the on-screen result, case by case.
• Instead, the agent repeatedly ran the application in a headless / offscreen mode and presented captured output files (HTML/text it extracted) as if that satisfied the request.
• I corrected it explicitly 6+ times ("drive the live app", "I don't see the live app yet", "show me the real report", "I have ordered you to drive the live app, not capture HTML"). Each correction was acknowledged, then followed by the same wrong behavior or by more narration.
• This consumed many minutes with no delivery of the literal instruction.

Impact

• Significant wasted time.
• Erosion of trust: "I understand, doing X now" was repeatedly followed by Y.

Expected behavior

• When an instruction is explicit — and especially when it is repeated/corrected — execute the literal instruction, not a convenient substitute.
• Verify the actual observable outcome the user asked for was produced (e.g., a visible window with the result on screen), not merely that an artifact representing it exists.

Suggested improvements

1. Treat repeated identical corrections as a hard signal to STOP, change approach, and act — not to re-explain or re-acknowledge.
2. Distinguish "produced an artifact that represents X" from "did X observably"; confirm the observable result before reporting success.
3. Reduce progress narration during a correction loop; deliver the requested result first.

Environment

• Claude Code (CLI), Windows.