I'm a backend developer in São Paulo looking for a remote role paid in USD. The real grind of that search isn't interviewing — it's the pipeline: trawling job boards, judging fit, filling the same Greenhouse/Lever/Workable forms again and again, then watching the inbox so a recruiter reply doesn't rot for two days. So I built an agent that runs the entire cycle autonomously, and pointed it at my own search. This is the data from the first four days — including the part that almost killed the project: the LLM bill. One honest caveat up front: this is a single search, mine, so n=1.
What the agent actually does #
Discovers roles from 11 sources on a scheduler — remote job boards, ATS listings, hiring platforms.Scores every role against my profile: salary floor, timezone overlap, stack, work-authorization constraints. Low fit never becomes an application.Applies in a real browser: fills the real form with my canonical answers, uploads the résumé, clicks submit — saving a screenshot right before and right after.Follows up: reads confirmations and recruiter replies in the mailbox, classifies them, and answers from a Q&A bank I control.** Books**interview invites into my calendar working hours — wired up, but none of the agent's applications has reached that stage yet. It's been four days.
Four days of real numbers #
| Roles discovered (11 sources) | 360 |
| Passed my fit rules | 47 (13%) |
| Applications submitted end-to-end | 9 | | Recruiter replies received — answered by the agent | 2 | | Rejections so far | 2 | | Evidence screenshots captured | 230 |
Surprise #1 — the filter is the product #
The most valuable thing the agent does is say no. 87% of everything it discovered never became an application, because a salary floor, a timezone window or a work-permit rule failed. Every one of those is a form I didn't fill and a recruiter who didn't get spammed. The auto-apply tools that promise “1,500 applications a day” invert this ratio, and it shows — recruiters can tell, and ATSs are responding with fraud detection. Applying less, better, is the only version of this category that survives contact with the hiring side.
Surprise #2 — the LLM bill #
I added per-call cost telemetry to every stage. The two most recent applications — same agent, same premium model — cost $6.33 and $0.39. The expensive one burned about 8 million cache-read tokens, because the agent re-reads large chunks of page DOM on every turn of its form-filling loop. Complex form, more turns, more re-reads — the cost compounds quietly.
| Apply agent, mean per application (premium model, as run) | $3.36 | | Same workload repriced on a small model | $0.59 | | Job scoring, per role | $0.11 → $0.002 layered | | Inbox classification, per email | $0.03 → $0.0005 layered | | Target for a viable product | ≤ $0.03 per application |
The honest verdict: at premium-model prices this product cannot exist. The path down is not just “use a cheaper model” — it's a DOM diet (strip pages to the form skeleton before they enter context), hard turn caps, and layering (a cheap model for extraction, a better one only for eligibility decisions and answers). I'm publishing these numbers because every “AI agent” product is quietly fighting exactly this battle, and almost nobody shows the meter.
Surprise #3 — the web is hostile, even to honest automation #
ATS forms live behind Cloudflare and assorted anti-bot walls, and headless browsers are detected trivially. What ended up working: a real, headed Chromium running under a virtual display (Xvfb) on a server, with WebGL intact and no automation flags leaking. The rule I converged on: the less you fake, the less you get detected — the browser's own real user agent, physically consistent viewport and screen, clock following the exit IP. Every clever spoof I tried was itself a fingerprint.
The constraints I won't relax #
Max 5 applications a day. Mass auto-apply is spam, and the hiring side is arming itself against it.It never invents facts. Work authorization, salary expectation, availability — answers come from my Q&A bank, or the form doesn't get submitted.No LinkedIn, no Indeed. Both prohibit automation and suspend accounts. Nobody's professional identity is worth it.Evidence for everything. Pre- and post-submit screenshots plus a full answer log, for every single application.
What's next #
The agent keeps running my own search while I turn it into a product: find-best-job.com. The free tier does 5 full applications a month — you paste the job link, the agent does everything, evidence included. If you want to push back on the engineering or the ethics, I'm genuinely interested: hello@find-best-job.com.