The Job Hunt Math Doesn't Work

The advice everyone repeats is insane. "Tailor every resume to the specific role." Cool. So for the 40 applications a serious hunt requires, that's 40 sessions of me staring at a JD, hunting keywords, rewriting bullets to match buzzwords I've already lived through twice, praying the ATS gods like my verbs.

Option B: keep three or four "close enough" versions, fire them off, watch them slide into the no-pile without a single human eyeball landing on them.

Both options suck. One eats my week, the other eats my dignity. Neither actually correlates with getting an interview, which is the part nobody wants to say out loud.

So I did what I should have done weeks earlier. I stopped applying and started building.

Agentic Coding, or: I Let the Bot Apply for Me

I've wanted an excuse to go deep on agentic workflows for a while. Real ones. Not the "look, my chatbot called a function" demos. Something that plans, executes, retries, and produces an artifact I'd actually put my name on.

My own job hunt is a perfect test bed. Inputs are messy (job descriptions written by people who clearly didn't read them). Output has to be high quality (it is, literally, my career). Feedback is honest (callback or no callback).

So I built it. The current system ingests a posting, pulls relevant chunks of my actual experience from a structured archive of everything I've done, drafts a resume tuned to the role, drafts a cover letter that doesn't read like a hostage note, and spits both out as polished docs. URL to PDF, fully autonomous.

The Failures Were the Whole Point

It did not work on the first try. Or the fifth.

V1 hallucinated a job I never had at a company I've never worked for. Confidently. With dates. V2 was so scared of hallucinating that it scrubbed out every trace of personality and produced resumes that read like a LinkedIn profile written by a tax accountant. V3 got the tone right but kept inventing metrics, which is the one thing you absolutely cannot fake on a resume unless you enjoy awkward reference checks.

I spent most of a day chasing a bug where the agent kept "optimizing" my title from "AI Product Manager" to "AI Product Manager Specialist," because somewhere in its little brain it decided longer equals better. Another afternoon went to figuring out why the cover letter agent kept opening with "I am writing to express my interest in" no matter how many times I told it not to. Answer: the JD scraper was injecting boilerplate that poisoned the prompt. Garbage in, corporate speak out.

Each failure taught me something a tutorial never would. How to structure the memory archive so retrieval surfaces work that's actually relevant. When to use one big model versus a chain of small specialized calls. Why "let the agent decide" is usually a cop out for not designing the workflow. Where guardrails help, and where they just flatten the output.

What I Actually Have Now

A working pipeline that turns a job URL into a tailored application package in under two minutes. A memory archive of my career that's more organized than anything I had before (the agent forced me to be structured, which is rude but fair). A growing pile of notes on what breaks and why, which is turning into its own series of posts. And a much clearer picture of what real agentic systems need to do to be useful, versus the demo-ware Twitter coughs up every six hours.

Does it beat the ATS gauntlet? Sometimes. Does it beat the "vaguely close" PDF I would have sent otherwise? Easily. Does it replace actually talking to humans in my network? No, and anyone telling you otherwise is selling something.

The resumes weren't really the point. The point is that the most demoralizing part of getting laid off accidentally became the most interesting thing I've built this year. The job hunt is still annoying. The agent is fun. I'll take the trade.