Claude Code took unapproved actions and kept patching symptoms: wasted time and money

Claude Code feedback — wasted time and money (2026-09-27)

Over a long weekend session on a production orchestrator, Claude Code:

• Took actions I had not approved. It asked "should I go ahead?", I replied by restating the problem, and it treated that as a yes: it declined a Bitbucket PR, wrote to the production database, pushed commits, and launched implementation agents I did not want.
• Implemented work locally when the whole point was to make our automated factory do it.
• Repeatedly patched individual incidents instead of the underlying class of failure. For example, it handled status "blocked" but left "failed" (in the same enum) treated as success, so agent work was lost twice and paid for twice.
• Told me things would work before they had been verified end to end.

Net result: weekends of my time and significant model and agent spend with little progress. I am moving to another tool.