Job-Hunting Agent Sent 160 Applications: LLM Fit Score Was the Worst Predictor of Replies
nivokvo · reddit · 2026-11-18
A math student in Belarus shares three months of logs from his job-applying agent: a Python daemon polls job boards, a cheap model filters, Claude Opus scores postings 0–10 and writes cover letters, a headless browser applies.
Key findings:
- The LLM fit score measures whether he'd like the job, not whether recruiters reply: AUC 0.48 on 121 outcomes.
- A small logistic regression with hand-set priors did better (0.67).
- The two engineering roles that replied were the ones the model ranked lowest.
- Guardrails beat prompts: every hard rule (no captchas, read full chat before replying, never apply without a letter) came from a real failure.
Code is open source with live numbers.
More from coding & agent
- Shipping a read-only MCP server to the Claude connectors directory: OAuth pitfalls and review notes — Fine_Airline_6832 · 2026-11-18
- Three Weeks Testing OpenAI's Astra: Strength Is Delegating Messy Stacked Context — round · 2026-11-18
- Unblocked cuts AI code review cost 3.1x running open models on Baseten — baseten · 2026-11-18
- "No SaaS, no apps": devs debate agents rewriting the entire software stack — alvelda · 2026-11-18
- AI-assisted coding is fueling an explosion of new music software — toptickcrypto · 2026-11-18
- Reflection AI launches Laser, an autonomous penetration-testing harness, plus a Trusted Access Program — shakoistsLog · 2026-11-18