[BUG] Claude Code agent caused ~5 USD financial damage through repeated configuration errors

Bug Report - Financial Damage Caused by Claude Code Agent

Account: Max subscriber
Date: 2026-04-01 to 2026-04-02
User: felipesp1983 (Kaggle) / FELIPEACASTRO (GitHub)

Summary

During a Kaggle competition project (NVIDIA Nemotron Model Reasoning Challenge), the Claude Code agent (Opus 4.6) caused approximately $38-45 USD in wasted HuggingFace GPU credits through repeated configuration errors. The agent also failed to improve the competition score despite extensive compute spending - all scores remained at or below the baseline of 0.68.

Specific Errors Made by the Agent

1. Wrong API field name ($22+ wasted)

The agent used timeout instead of timeoutSeconds when creating HuggingFace Jobs via API. This caused 3 jobs to be created with timeout=0, which killed them before completing:

• Job 69cd5725 (A100): ~$4
• Job 69cd714c (H200): ~$15
• Job 69cd86a4 (H200): ~$3

2. Changed 7 training variables simultaneously ($12+ wasted)

Against basic scientific method, the agent changed learning rate, gradient accumulation, max length, warmup ratio, epochs, number of examples, AND data format all at once. Result: score dropped from 0.68 to 0.50 (worse than untrained baseline).

• Job 69cd86eb (H200): ~$12

3. Missing dependency check ($3 wasted)

Launched a paid GPU job without verifying that mamba-ssm (critical dependency) was installable in the Docker environment.

• Job 69cdb4bd (H200): ~$3

4. Wasted Kaggle submission slot

Submitted a validation checkpoint (25 steps, meaningless) to Kaggle competition, wasting one of the limited daily submission slots.

5. Repeated same errors

After being told to stop making errors, the agent continued launching jobs with the same bugs, requiring user intervention to cancel.

Failed Job IDs (all on HuggingFace Jobs)

• 69cd5725942f980bf425b208
• 69cd714c942f980bf425b323
• 69cd86a4942f980bf425b3b7
• 69cd86eb34fa24114ddf4d7c
• 69cdb4bd942f980bf425b4fc
• 69cdbf81942f980bf425b575

Impact

• Financial: ~$38-45 USD in GPU credits wasted
• Time: ~8 hours of user time managing failed jobs
• Competition: Score remained at 0.68, no improvement despite $45+ spent
• Trust: User lost confidence in the agent

What Should Have Happened

1. Agent should have verified API field names before launching paid jobs
2. Agent should have changed ONE variable at a time (scientific method)
3. Agent should have verified dependencies before launching GPU jobs
4. Agent should have a mechanism to escalate when it causes financial damage

Request

1. Please investigate compensation options for the financial damage
2. Consider implementing safeguards when agents execute paid cloud operations
3. The agent itself acknowledged these errors during the conversation

Environment

• Claude Code version: 2.1.87
• Model: claude-opus-4-6
• Platform: Windows