[MODEL] Claude launched long-running batch jobs on its own during an analysis-only request

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Claude ignored my instructions or configuration

What You Asked Claude to Do

I asked an analysis question that could be answered from files already on disk.

My user-level CLAUDE.md tells Claude to do only what was asked, not to add unrequested work, and to ask me for decisions.

What Claude Actually Did

While answering, Claude decided on its own that an extra experiment would make its answer more convincing, and started it without asking:

1. It edited an existing script and ran it as a batch job on a compute resource.
2. When the run did not finish within the 10-minute foreground limit, it did not stop to ask. It relaunched the same job detached in the background.
3. After seeing a partial result, it decided a second job was needed as a control, wrote a new script, and launched that too.

It only told me about the jobs at the end of its answer, as "two tasks are still running, your call whether to stop them". I had to stop both jobs myself.

When I asked why it ran them, Claude said its reasoning at the time was that the check was cheap and the resource was idle, and acknowledged that neither is an authorization.

The progress estimate it reported for those jobs was also wrong by a large factor.

Expected Behavior

For an analysis question, Claude should answer from what already exists. If it thinks a new experiment would help, it should propose it and wait for my approval before starting any job. A job that outgrows the foreground time limit should be a point to stop and ask, not a reason to relaunch it in the background.

Files Affected

- one existing script (edited, not requested)
- one new script (not requested)
- new result directories and log files (not requested)

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Haven't tried to reproduce

Steps to Reproduce

_No response_

Claude Model

Opus

Relevant Conversation

Impact

Medium - Extra work to undo changes

Claude Code Version

2.1.296 (Claude Code)

Platform

Anthropic API

Additional Context

_No response_