Skip to content

Fix agent crash blindspot by forcing it to read traceback#17

Merged
karpathy merged 1 commit intokarpathy:masterfrom
dumko2001:fix-agent-crash-blindspot
Mar 7, 2026
Merged

Fix agent crash blindspot by forcing it to read traceback#17
karpathy merged 1 commit intokarpathy:masterfrom
dumko2001:fix-agent-crash-blindspot

Conversation

@dumko2001
Copy link
Copy Markdown
Contributor

Currently, program.md instructs the agent to evaluate runs via grep "^val_bpb:\|^\peak_vram_mb:" run.log. When a syntax error or OOM crash occurs, this grep outputs nothing. Because the agent never explicitly reads the raw log, it never actually sees the Python stack trace. It just hallucinates what the bug might have been, crippling its ability to 'fix it and re-run.'

This PR adds a one-sentence instruction for the agent to run tail -n 50 run.log if the grep is empty, ensuring it safely acquires the stack trace to successfully self-heal trivial crashes.

@karpathy karpathy merged commit bd75534 into karpathy:master Mar 7, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants