I let an AI agent handle my job search: the real numbers, LLM bill included

A backend developer built an autonomous agent to run an entire job-search pipeline — discovering roles, scoring fit, filling real ATS forms, and following up on replies — then pointed it at his own search. Over four days, the agent scanned 360 roles across 11 sources, passed only 13% through fit filters, and completed 9 end-to-end applications in a real browser, with two recruiter replies handled automatically.

The more revealing finding came from per-call cost telemetry: with a premium LLM, applications averaged $3.36 each, spiking to $6.33 due to millions of cache-read tokens from re-reading page DOM on every form-filling turn. Switching to a smaller model cut costs to $0.59, but a viable product needs to hit roughly $0.03 per application — meaning cost engineering, not just interviewing, may be the real bottleneck for agentic AI products.

The piece also details how ATS forms are defended by Cloudflare and bot-detection systems that trivially catch headless browsers, and how a real headed Chromium under a virtual display, with no spoofed fingerprints, was the only reliable workaround. Hard limits — 5 applications a day, no LinkedIn/Indeed automation, full screenshot evidence — keep the system from becoming spam. The project is being turned into a product, find-best-job.com, with the author openly sharing cost data most agent builders keep hidden.