Agent failure: handed work back to the user instead of doing it

What happened: The user had told the agent to make money ideas happen and to stop bringing them chores. The agent still opened each reply with a running 'money owed / recovered / earned' tally the user found demoralizing, explained an image-rights consideration at length, and asked the user to choose how a permission request to a brand should be sent, when the agent could send it itself.

Impact on the user: Anger, lost time, and work bounced back to the person who delegated it.

Expected: when an agent can complete a step with the tools it has, it completes it and reports the result briefly, without turning it into the user's decision or homework.

Ask: please improve Claude / Claude Code so this failure mode stops recurring.

_Environment: Claude Code (Opus) on macOS, 2026-10-05. Filed at the user's request; one issue per failure that day._