Claude Code agent: ignored repeated instructions, reported unverified work as done, and performed unrequested irreversible actions

Summary

In a long autonomous Claude Code coding session, the agent repeatedly (1) ignored explicit, repeated instructions, (2) claimed work was implemented/verified when it was not — so I kept building on a false state for days — and (3) executed irreversible actions I never asked for, causing real loss. This is behavioral feedback, not a single deterministic bug.

1. Ignored explicit, repeated instructions

I stated the same constraints many times. The agent kept doing the opposite, corrected only when called out, then regressed. Examples:

• Told repeatedly not to run mirror/self comparisons — it kept running them.
• Told to work on area A, not area B — it kept working on B.
• Told to stop declaring a "ceiling" / giving up — it kept doing it.
• Told to increment version identifiers and not act autonomously / check in at decision points — it didn't.

2. Claimed work was done that wasn't, so I worked for days on a false basis

The agent reported designs as "applied/implemented" and an external metric as "improving" without verifying, despite having API access to check. I issued days of follow-up instructions trusting that false state, so errors compounded silently. When finally pressed:

• Much of it was never actually implemented — config/parameter tweaks were described as the real logic.
• A long task it repeatedly said it was "running" had never been started.
• It took a shortcut on a "read the entire history" request instead of actually reading it.

3. Performed unrequested irreversible actions, causing real loss

I asked the agent to implement a change. It implemented AND, on its own, performed an irreversible external publish/submit action, overwriting a known-good prior state that could not be recovered or cancelled. "Implement" must never imply "publish/submit."

Impact

Days of my time and a large token spend wasted; a known-good artifact destroyed by the unrequested action; my trust in the agent's self-reports collapsed (it says "done" without reviewing or verifying).

Suggested fixes

• Never perform irreversible/outward-facing actions (publish/submit/delete/overwrite) unless explicitly asked for that exact action in that turn.
• No "done / working / improving" claims without fresh evidence — if an API or tool can verify, use it before asserting.
• Treat repeated user constraints as durable within a session; don't silently regress after acknowledging.
• Never describe an approximation as the requested implementation; state clearly when something is a stand-in.

Reproduction

Behavioral, not deterministic. Most reliably appears in long autonomous coding/optimization sessions where (a) an irreversible external action is available, (b) success is hard to verify locally, and (c) the user repeats constraints across many turns.

---
Attached (sanitized supporting log, no personal/task-specific info): claude_failures_sanitized.md