How to Put AI's Hands on Your Keyboard with Computer Use Agents

What computer use is, what it's good for, and the 15-minute move: turning your Monday dashboards into one briefing an agent compiles for you

So, here's what my Monday mornings used to look like. Seven newsletters on Beehiiv, which means seven analytics pages. Two sites on Vercel — theAIE.net and AllThingsai.org — which means two web analytics dashboards. Plus social. I'd open five tabs, copy a number, alt-tab to a doc, paste, alt-tab back, repeat. Forty-five minutes of being a human clipboard before my first coffee was cold.

I've been doing some version of this for thirty years. My first tech job was ISP support in the 90s, and one of my duties was compiling the morning server report — walk the numbers from three terminal screens into one page so the boss could read it over breakfast. In 1996 that was a job. In 2026 I was still doing it, just with prettier dashboards.

Not anymore. Now an agent walks the tabs and I read one table.

// The Takeaway: Computer use is the class of AI agents that operate software the way you do — they take screenshots, read what's on screen, move the mouse, click buttons, and type. Chat AI gives you advice about your dashboard; a computer use agent works in your dashboard. It's built for exactly the work that has no API you'll ever call: admin consoles, analytics pages, settings panels, the fifteen browser tabs your job actually lives in — and, with Claude's desktop agent, your local files too. The move this week: pick the dashboards you check every Monday and hand the collecting to an agent.

Type /agent in ChatGPT or open Claude in Chrome → ~15 minutes · Needs a paid Claude plan or ChatGPT Plus/Pro (40 agent messages/month on Plus, 400 on Pro)

For three years, "using AI" has mostly meant a chat window. You describe your problem, the AI describes a solution, and then you go do the clicking. It's like calling tech support in 1997: helpful, but the guy on the phone can't see your screen, so half the call is you reading things off the monitor to each other. I know. I was the guy on the phone.

Computer use closes that gap. Anthropic's computer use tool runs a simple loop: the model takes a screenshot, decides what to do, sends a click or a keystroke, takes another screenshot, and repeats until the job is done. That's it. No integration project, no connector, no waiting for a vendor to ship an API. If a human can operate it with a mouse and a keyboard, an agent can too.

Where it shines: dashboards, admin consoles, legacy web apps, and especially cross-tab work — "pull this from the analytics page, compare it against the spreadsheet, put the summary in the doc." That's the work nobody ever automated because no single tool owned it. Where it doesn't: fast real-time work (agents click slower than you do), and anything high-stakes and irreversible. An agent is a very fast first-week intern. Give it intern work, not the keys to billing.

// The real shift: Integrating software used to mean APIs, engineers, and a quarter of somebody's roadmap. Now the integration surface is "has a screen." Every dashboard your company pays for — the ones IT never budgeted to connect to anything — just became automatable by the person who actually uses it. That's not a product feature. That's a change in who gets to automate.

Claude in Chrome is Anthropic's browser agent — it lives in a Chrome side panel and works in your browser, with your logins, which makes it the natural pick for dashboard work. It launched as a research preview for 1,000 Max subscribers in 2025; it's since opened to all paid plans (Pro, Max, Team, Enterprise) via the Chrome Web Store. It asks permission site by site, blocks itself from financial sites, and Anthropic published honest numbers: safeguards cut prompt-injection attack success from 23.6% to 11.2% in testing. Not zero. Remember that.

ChatGPT agent mode runs in a virtual browser on OpenAI's side — type /agent in the composer (or pick it from the tools menu) and it goes off and works. Plus gets 40 agent messages a month, Pro gets 400. Because it's not your browser, logging into your dashboards means its takeover mode: the agent pauses, you type credentials yourself, and it can't see them. A little clunkier for logged-in dashboard work, but it does the job.

The computer use API is the developer route — a beta tool where Claude drives a full desktop (screenshots, mouse, keyboard) inside a sandbox you run. This is how you'd build the repeatable, scheduled version for your team. Not writing code this week? Skip it — but know it exists, because it's what the other two are built on. And if the work lives in files instead of tabs, Claude Cowork — Anthropic's desktop agent, launched in January and now on web and mobile — works directly in your local folders and browser. Same idea, pointed at your hard drive.

This is my actual Monday workflow. Mine is Beehiiv, Vercel analytics, and social — swap in whatever you check weekly: newsletter stats, web analytics, CRM pipeline, ad spend.

Total agent time: about ten minutes. Total Mark time: reading one table and noticing the things worth noticing — which was supposed to be the job all along.

Once the collect-and-compile pattern clicks, you'll see it everywhere. Point the agent at three competitors' pricing pages and your notes doc: "Extract plans, prices, and limits into a comparison table." Or at that dashboard-to-spreadsheet chore everyone has: "Read the numbers off this admin console and fill in columns B through F of this sheet." The skill isn't Beehiiv or Vercel. The skill is writing a work order a junior tech could execute — named sources, exact fields, clear boundaries. I learned to write those dispatching actual junior techs in 1996. Turns out it transfers.

Computer use spent 2025 as a developer beta and a 1,000-person pilot. This year it landed in mainstream plans — any paid Claude plan gets the Chrome agent, any ChatGPT Plus subscriber gets agent mode, and Cowork put the same idea on the desktop. The constraint isn't access anymore. It's whether you've learned to write the work order. The people practicing on low-stakes chores now — briefings, tables, extractions — are the ones who stop being the bottleneck when the chores get bigger.

One caveat, and I mean it: start read-only. Keep agents out of billing, account settings, and anything you can't undo — even Anthropic blocks its own agent from financial sites, and they built the thing. Extraction first. Trust is earned, even by software.

See you Wednesday with the Advantage.