A Fully Automated AI Job Application Agent

To be honest: I’m not entirely comfortable with this. I can’t quite put it into words; it feels somehow wrong — yet at the same time quite exciting. No, this isn’t about infidelity — I’m too honest for that. It’s about AI-assisted job applications. But this isn’t yet another article about the “ultimate prompt for the perfect cover letter.” I’ve gone quite a bit further — right up to the edge of what’s possible.

I have a digital butler I’ve aptly named “Carson” — after the character from Downton Abbey. He now plays a small but central role: he searches for jobs that match my profile, researches everything about the company and position, develops an application strategy, writes the cover letter, submits the application, and then logs the whole thing with the Regional Employment Centre (RAV) as evidence of my job-seeking efforts. Yes, the complete package — (almost) autonomously, so that I could prune the apple tree in my allotment garden in the meantime.

How Curiosity Turned Serious

Why did I actually do this? The simple answer: because I can. Honestly, I didn’t plan it this way — there’s that infidelity analogy again, which I don’t think is entirely off the mark. It all started when I was made redundant and got completely drawn into the whole AI thing.

As often happens with a neurodivergent brain like mine, you just start experimenting wildly. What’s fascinating about AI is: even when you can’t imagine in your wildest dreams that something could work, you just try it and find out — hey, that actually works! So first I put together a framework that drafted tailored cover letter suggestions. Then came a letter template in LaTeX, and suddenly I had a foundation. Early in the year, Peter Steinberger appeared with his AI agent. I kept experimenting and stitched everything together somehow, until I had a system I’m not sure whether I should be proud of — or whether it simply “just happened.”

A Consultant for Self-Marketing

Another reason is certainly that I’m not particularly good at ingratiating myself with people. Self-promotion just isn’t in my nature, and I’m only on LinkedIn because you have to be these days. The AI can market me better than I can do it myself — so why shouldn’t it? The headhunter who placed me in my last job essentially did the same thing and charged a commission for it. Carson makes do with an API budget more modest than my monthly coffee spending.

Levelling the Playing Field in the Algorithm Jungle

Ultimately, my project is also an expression of systemic critique. Anyone applying for jobs right now — particularly in fields with high volumes of applicants — is frequently filtered out automatically. Rejections sometimes arrive so quickly that no human could possibly have read the application — if a response comes at all. Often you receive a superficial rejection that you simply can’t do anything with. These are text blocks so generically worded that any halfway attentive person can tell: the same software was used here.

It’s often no secret that recruiting teams receive so many applications that they can’t study them in detail or provide individual feedback. That’s understandable to a point. When LinkedIn now has an “Easy Apply” button for fullstack developer jobs, plenty of people click it and count it as a job-seeking effort — even if the only web project they’ve ever done was a site for their uncle. Or the barista who imagines himself as a “vibecoder” with a golden future in Silicon Valley.

So why shouldn’t I make full use of my arsenal of possibilities and upgrade technologically? If the other side is using algorithms, I’ll send my butler Carson into the ring.

A Look Under the Hood

Now to the technical side, because many people who aren’t closely following AI developments wonder how something like this is actually built.

The operating system of the whole thing, so to speak, is an AI agent based on NanoClaw. Similar to OpenClaw, which is currently generating so much hype, it’s nothing more than a process that continuously searches for tasks — in this case in a SQLite database. Tasks can be simple messages from channels like Telegram or WhatsApp, or recurring defined prompts such as the daily job search or a goodnight joke — though most models simply fail at inventing jokes. Perhaps I should become a comedian; that seems to be one of the few professions not yet under threat. Anyway, these tasks are processed using the full infrastructure of a modern operating system. NanoClaw launches isolated OS containers to keep different areas separate, so no major damage can occur and nothing accidentally ends up in the wrong channel. There’s a family group with a WhatsApp channel, completely separated from my private matters and job search, which is handled via Telegram.

If you’re picturing me sitting at a desk with multiple monitors, entering glowing green command-line instructions, writing programs, and analysing streams of data — you’ve watched too many mainstream sci-fi films. I talk to a Telegram bot as if it were a person: “Could you please prepare the application for position XY?”, “Please log the job-seeking efforts for XY in Job-Room” or, somewhat more casually: “Keep going, no rest now! I’ll grab my phone to scan the QR code.” These are real messages taken from the chat history. The trick is communicating with these systems exactly as you would with real teammates. You get used to it quickly. What remains more unsettling is the moment you realise that the machine communicates more politely than the last recruiter who sent you a rejection.

In the background, NanoClaw starts a container for my application group via Docker and mounts a group folder from the host. This folder contains two repositories: one is the framework mentioned above for creating cover letters, the other is a tool for generating CVs using LaTeX, including copies of all PDFs of employment references and certificates. When I write a message like “Submit the application documents,” it’s passed via inter-process communication (IPC) to the container where the task is executed by Claude Code. Once complete, the host sends the result back to the Telegram group.

The magic happens inside the container, which is nothing more than a temporary, stripped-down Linux operating system. All the applications and tools that a regular user would have access to are installed there — for example, a browser, email client, or calendar. Claude Code can operate these to a certain extent. Often it’s not even necessary, because if we’re already talking about magic: Claude Code can be described as a wizard that’s extremely good at generating code and simply writes the tools it needs itself. For instance, Carson put together a small Python script to send emails with attachments. A separate email programme didn’t need to be installed at all.

The Instructions

What at first glance sounds like complex automation reduces, in practice, to clear, sequential instructions — not much different from a checklist you’d hand to a new team member:

• Job search: Based on my profile, a deep search is initiated via Parallel AI. Additionally, three Swiss job portals are searched for suitable positions. The agent keeps an internal log of which positions have already been presented to me, which ones I’ve applied for, and the current status. New positions are suggested to me daily in order of priority.
• Preparation: If I’m interested in a position, I instruct the agent to prepare my application documents. The cover letter framework has built-in commands that the agent works through in sequence.
• Research: A further deep search is conducted on the company and the advertised position. All artefacts such as copies of the job advertisements, search results, and my personal notes are saved per application.
• Strategy & text: Based on this information and my personal job profile, an application strategy is developed. From this, a cover letter is written whose phrasing is precisely aligned with the strategy, to convey my suitability and motivation accurately and authentically.
• Review & submission: I check the result and suggest adjustments if necessary. When everything fits, I have the application submitted. This happens by email, web form, or job portal, depending on requirements.
• Evidence of job-seeking efforts: The agent logs into the Regional Employment Centre (RAV) via browser and records the job-seeking efforts in Job-Room. This is one of the steps where I still need to personally scan a QR code. Welcome to 2026: the human has officially been reduced to the function of an authenticator.

What’s fascinating is that neither a technical standard nor a cryptic programming language is needed to instruct the agent. They are ultimately just text instructions that any person would understand and could follow — just not as quickly as a machine. Here’s an excerpt from the instructions: sample_application_workflow_de.md

Utopia, Dystopia, or Simply the New Reality?

It feels strange — a bit like a psychedelic trip, or a dream I’ll wake from soon. Whether this is closer to utopia or dystopia, I haven’t quite figured out for myself. What bothers me most about this experiment is the contrast in how AI as a technology is perceived among professionals. In expert circles, I keep encountering people who are very sceptical, who are still “observing” or dismissing it as hype. How I wish I could join that chorus! I genuinely keep trying to question the whole thing. I often hope the big AI bubble will burst and I can go back to working in good conscience as a reasonably capable software developer until retirement.

But experiences like this job application agent show a completely different side — one that escapes my comprehension and cognitive capacity. And then I’m always left with the big question: why do critics downplay the potential so much? It’s not some theoretical future; it’s simply the hard reality that you probably just have to experience firsthand to accept it, at least in part. Because yes — with the effort of perhaps four working days and the budget of a decent dinner out, I’ve automated the tedious application process in a fairly agile and unplanned way, so that I’m now just the “human-in-the-loop.”

Whether this is utopia or dystopia, I haven’t decided — and I’m not sure the question even needs an answer. What I know is: it works. And it does so before anyone in any expert circle has finished debating whether it should.

Carson, for his part, has no opinion on the matter. Good butlers don’t.