Vol. I · No. 579Rochester, New York, Thursday, September 24, 2026Browser Agents · Page B9
All workSection B · Page B9

Browser AgentsJobOps

The Job Hunt Agent That Refuses to Hit Submit

JobOps finds jobs, tailors a resume for each one and pre-fills the application, then stops and waits for me.

Graphic6 steps
  1. DiscoverListing repos, HN and a Greenhouse, Lever and Ashby crawl
  2. ClassifyDeterministic region and start-date rules
  3. ScoreHeadless Claude rates each job 0 to 5
  4. TailorLaTeX resume and cover letter per job, with a change log
  5. Pre-fillGuarded browser agent fills the form in Chrome
  6. ReviewI check everything and submit it myself
What happens inside JobOps, in order.Graphic: The Sahil Bachu

Applying to new-grad roles is a second job. JobOps is the tool I built for my own search. It finds postings, scores them, writes a tailored resume and cover letter for each one, and pre-fills the application form. Then it stops. A human, me, reviews every application and clicks submit.

Discovery pulls from five sources: two GitHub new-grad listing repos, Hacker News Who is Hiring, a direct crawl of Greenhouse, Lever and Ashby job boards, and one best-effort job site. If one source breaks, the others keep going. The board list grows by itself: any board URL seen anywhere gets added, and dead boards get switched off. The local database now holds over 25,000 jobs, and the board index has passed 900 boards.

The rules live in code, not in the prompt, so the model can't talk its way past them.

Before any AI touches a job, deterministic classifiers handle the easy calls. A tiered region resolver keeps Vancouver, WA and drops Vancouver, BC. A start-date ranker prefers January starts, since I graduate in December 2026. Then Claude Code in headless mode scores the rest in batches of 25 against my profile. The output has to be strict JSON that passes validation, with one retry if it doesn't parse.

For each job I pick, it rewrites my master LaTeX resume and cover letter under a no-fabrication rule, compiles them with a bundled tectonic binary (with up to 3 fix-and-retry attempts), and writes a CHANGES.md so I can see exactly what moved. A voice linter rejects answers with dashes or about 40 AI-sounding phrases and makes the agent rewrite them.

The autofill agent is where the guardrails matter. It's a sandboxed Claude whose only tool is a 316-line guard script. The guard refuses to click submit, apply or send. It refuses sign-ins, account creation and consent boxes. It only navigates to that job's form and only uploads that job's PDFs. Page text is treated as untrusted data, and anything that looks like a prompt injection gets reported. The rules live in code, not in the prompt, so the model can't talk its way past them.

To prove it holds, there's a mock application site with normal and deliberately hard forms, and a self-test that only passes if nothing got submitted. Overnight mode picks the top 10 jobs, tailors each one and leaves every form pre-filled in its own Chrome window, waiting for me in the morning.

More from The Work

See all 25

Generative UI

A Canvas of Widgets You Order in Plain English

An infinite canvas where an AI agent turns a plain-English request into a live widget wired to your own data.

Ask for something like a card that summarizes your unread emails, and an agent researches, designs and places a live widget on an infinite, zoomable grid. Connectors for Gmail, Calendar, Slack, GitHub, Linear and Notion feed it real data.

Continued on Page B11

Agentic AI

Your Coding Agent Fills the Gaps in Your Resume

A local-first orchestrator that reads a student's resume, proposes projects to fill the gaps and has their own Claude Code or Codex build and deploy them.

ResumeGod reads your resume, suggests projects that would fill its gaps, then drives your own Claude Code or Codex to build and deploy them. It finishes with a personal site and a fresh LaTeX resume. Nothing is hosted centrally.

Continued on Page B15

Multi-agent

Hire AI Employees, Then Walk Over to Their Desks

A pixel-art office where every AI agent is an employee with its own desk, skills, scheduled jobs and Docker container.

PixelHQ turns the open-source SkyOffice into a single-player office for AI agents. Walk up to an employee, chat with it, give it skills and scheduled jobs, and feed a shared knowledge base. Each employee runs in its own container.

Continued on Page B17