AI Update, June 2026: From Helper to Operator

*Published on Senior Dev Notes and LinkedIn, June 30, 2026.*

A few months back I wrote that AI had changed my career. That post was about pair-programming with Claude, autocomplete, getting unstuck faster. All true, all still true. But the stack underneath my day has moved past that and I want to write down what it actually looks like now, because what I read about AI online still mostly describes the 2024 version of the job.

The short version: I stopped using AI as a helper and started using it as an operator. It opens browsers, signs in to my accounts, edits config files on remote servers, creates API keys in dashboards I haven't touched in months, files Freelancer bids while I'm cooking dinner, and patches itself when things break. I sit a level up from that and decide what should happen next.

Here is what a week actually looks like.

The browser is just an interface to other people's APIs now

Half of what I do involves clicking around on a third party's website. GitHub for creating an App and minting installation tokens, Google Cloud Console for a new OAuth client and downloading the JSON, Facebook Marketplace for posting a listing or pulling lead data, an ATS like Greenhouse or Workday or SuccessFactors for submitting a job application, Twilio for spinning up a sub-account, GoDaddy for adding a DNS record. None of these have decent APIs for the things I want. Some have APIs that require an API key that you can only get from the dashboard, which is the funniest catch-22 in software.

So I drive the browser. A noVNC sandbox in a container, a Playwright shell that takes screenshots back to me, a Chrome extension for the cases where the page detects automation (Google's Cloud Console is the worst about this, the "Download JSON" button literally will not respond to a synthetic click event). Claude reads each screenshot, decides where to click, types into fields, sometimes I lean in and solve a captcha by hand in the noVNC window. The work is split between the model and me on a single page, in real time, and it feels different from anything I've done before.

The thing I want to convey is that I do not micromanage this anymore. I tell Claude "make a new Google OAuth client called watchnotify, restrict it to these scopes, download the JSON, drop it at this path on this server" and walk away. It comes back with the JSON in place and a summary of where it stumbled. Sometimes it stumbled at a checkbox that Patchright cannot tick (a known issue with Google's sensitive-scope consent screen) and I open the noVNC, tick it, and tell Claude to resume.

One Claude talking to another Claude (or to me, the copy/paste pipe)

The interesting workflow is when I run two or three sessions at once. One is doing research, reading code, building a plan. Another is doing the implementation, with edit and bash permissions. A third is driving a browser. They don't talk to each other directly. I'm the copy-paste layer between them, which sounds dumb until you try it. The research session writes a 200-line plan, I drop the plan into the implementation session, the implementation session writes a different 200 lines back about which files it touched, I drop that into a third session that writes the test plan. It is faster than doing any single one of those steps myself by a lot, and the sessions are short enough that each one stays sharp.

When I want it fully automated I use claude -p (one-shot prompts). I have shell scripts that run claude in a loop, each iteration feeding it a fresh problem and a list of tools it can use. The auto-bidder on my gigfinder app uses this to draft cover letters for each new Freelancer project (it pulls the project description, decides if it's worth bidding, calls Claude to draft, posts the bid). It also updates its own playbook file when it learns something new about a particular client or platform, so the next run is smarter. That self-updating-playbook loop is the part I would not have believed in 2024.

Images, videos, and creative output

I generate hero images and short product videos with Higgsfield (their CLI is great if you keep auth refreshed, my cron rotates the token every 30 minutes because their session expires aggressively). For longer cuts I stitch Sora 2 clips together with a Remotion pipeline. The DTF Paper Co site has a "printer at work" hero loop that took me about 25 minutes from "I want a printer doing print things" to a deployed MP4. A traditional shoot would've been a half-day plus location. I am very aware that this puts a lot of stock-footage and small-shoot work in a weird spot.

I am also writing more. Not because I love writing (you can see I don't) but because content for SEO is now cheap to produce and the bottleneck is having opinions worth publishing. I generate a draft from a research session, edit it down to half the length, add the parts only I would know, publish. The drafts have a tell to them. The editing is the work.

Job hunting at scale

I built an automated apply pipeline for myself. It pulls postings from LinkedIn, Indeed, and direct ATS scrapes, filters out the ones that want Smartapply or Easy Apply (low signal-to-noise), and submits to the real company ATS with a cover letter tailored to the posting (separate tone for tech vs non-tech roles, separate resume cluster depending on what they're hiring for). It's not a magic "get a job" button, but I do not spend an hour writing each cover letter anymore and that means I apply to ten times more roles. The ones that bite, I respond to in person.

Ad campaigns and the boring back-office stuff

I run ads for a couple of my own things and for two clients. Google Ads and Meta. I have Claude open the dashboard, pull the last 7 days of metrics into a CSV, identify which ad creatives are underperforming, write replacement copy in the brand's voice, and stage the new ads as drafts for me to approve. Approval is the human gate. Everything else is automated.

Same pattern for SEO. Keyword research, on-page audits, schema markup, internal linking maps, content briefs that go to a writer (sometimes another Claude session). I have a Common Crawl domain-rank lookup service running on a $5 VPS that any of my projects can hit to figure out if a backlink target is worth pursuing. None of this is novel individually. The shift is that all of it runs as background tasks while I sleep.

Troubleshooting local computers

This is the one I would have called a parlor trick a year ago. A friend's Windows machine started bluescreening on boot. I had her open a remote support tunnel, Claude read the minidump file, identified a specific Realtek driver, walked her through rolling it back from Device Manager, fixed in 12 minutes. I have done this with corrupted registry entries (Claude diffs against a known-good export), memory leaks (it reads procmon traces), and missing DLLs. The same kind of help an experienced sysadmin friend used to give over the phone, except it is faster and it costs me $0 and they don't have to wait for me to be free.

The bad parts: outages and rate limits

I do not want to leave the impression that this is frictionless. Claude Code has periodic outages, sometimes for an hour, sometimes for most of a workday. When the API is down, none of this works. The Anthropic status page becomes the most-refreshed tab in my browser. I have rate limits on heavier days that pause my agents until the window resets. I keep three Max accounts warm via a token-refresh service I wrote (an unattended host pulls a fresh credentials file every two hours so the cron-driven scripts never have to interactively re-auth) and I rotate between them when one hits its cap. Not elegant, but works.

The other failure mode is more subtle. The model gets confident about something that is wrong, executes the action, and breaks things in a way I have to manually unwind. I have lost work to this. I am more careful with destructive actions now (rm, git push --force, anything touching production data) than I was in March. The rule I landed on is: anything reversible, let it run; anything not, confirm.

What I think this means

I don't have a tidy thesis. I'm a senior full-stack engineer who used to write code all day and now spends most of my day designing pipelines, supervising agents, doing the small share of work that needs my judgment, and writing more than I ever have. My output as a working unit has gone up by something like 5-10x depending on the week. My job title has not changed but the description should. I get a lot more done. I also get sloppier when I am tired, because the cost of asking Claude to do something is so low that I sometimes ask before I have thought.

If you are still using AI as autocomplete, you are missing the bigger thing. The bigger thing is that the agents will go do the work, you just have to know what work to ask for.

-Dino

---

*If you want to put AI agents to work on your own projects (lead pipelines, ad ops, content systems, browser-driven integrations, SEO at scale), I help individuals and small teams design and run setups like the ones above. Learn more at helpwithweb.com.*

*- Dino Bartolome*