Shut Those Laptops! Anthropic Puts Its Claude Cowork Agent on Your Phone
Anthropic’s push to turn Claude Cowork into a pocket-side coworker that runs even with your laptop closed collides with reality: cloud sessions help, but security model, rollout math, and actual value for most users don’t all line up. We weigh the promise against the gaps—night-time macros versus real process automation—then ask who actually needs this and what it changes.
Transcript
Laura —okay, so the headline is ‘close the lid, the agent lives’ — that’s the sell?
Harper Nah, it’s worse — it’s ‘your always-on agent that never sleeps, even when you do’, and the demo sells the fantasy, not the real gap.
Laura Sure, but hold on — it actually keeps working when the laptop’s shut? That’s not nothing. You used to have to prop the thing open like a laptop sacrifice to Moloch.
Harper It does — but what it keeps working on is the list of tasks Anthropic pre-approved and blessed with a timeout budget. That’s not… a general agent.
Laura Wait— the Wired piece shows a renewal doc assembled overnight from Slack threads and email. That’s real output, not just ‘agent polled API and returned JSON’
Harper Right, right — but assembling a shopping list from scattered inputs is not the same as authoring a negotiation-ready document. The agent is doing the data janitor work, not the persuasion work.
Laura Okay, fair — but if the janitor is freeing me to do the persuasion, that’s a product win. You’re acting like this is all trickle-down Silicon Valley puffery.
Harper It’s not puffery; it’s scope creep. They rolled ‘run in the background’ into ‘take over parts of my workflow I used to do myself’ without proving it’s reliable.
Laura —okay, fair. But the article says usage is real: white-collar tasks like data reports and slide decks are skyrocketing inside Max customers. That’s not vaporware chatter.
Harper Data reports and slide decks are the no-brainer checklist tasks that every tool promises. The article’s own footnote admits prompt-injection risk is still real — so what you’ve really got is a macro runner with a cute agent skin.
Laura That’s reductive — the demo showed a doc crafted from scratch, not a templated slide deck. You’re dismissing structure and voice because you’d rather write it yourself.
Harper I’m dismissing because ‘write a reference doc’ is still mostly ‘search our corpus, stitch paragraphs, apply tone policy, then hope it reads like a human.’ The stitching is where it leaks.
Laura Then explain why usage numbers are real and climbing. Anthropic’s report says ‘business process and operations’ is now a top category — like actual workflows, not just toy tasks.
Harper Because Max subscribers are the ones who have the budget to tolerate the quirks. For everyone else, a 24/7 agent is overkill unless your job is literally gluing together reports at 2 a.m.
Laura Not everyone — but the wedge isn’t the overnight chores, it’s the phone-as-controller. That’s the real unlock: I can queue tasks when I’m not at my desk, and the laptop can stay shut.
Harper The phone-as-controller existed via Dispatch — the novelty is the background runtime. And background runtime on someone else’s VM is a risk surface they haven’t fully locked down.
Laura They added session isolation and tighter timeouts. The article says it’s beta — so yeah, it’s not perfect, but it’s a direction.
Harper Direction that costs a hundred bucks a month to try. For most people, rolling a cron job and a Slackbot is still cheaper and less terrifying than handing their renewal package to an always-on cloud VM.
Laura —right, that’s what I meant. But if you’re a Max user who juggles async tasks across devices, this changes the game. You’re paying for continuity, not just features.
Harper Only if the continuity doesn’t evaporate when Silicon Valley pivots to the next crustacean meme. OpenClaw started this trend, and now every Big Tech agent is chasing background runs — but which one actually ships a story users trust?
Laura Anthropic at least has a ship story — the article’s own reviewer tried Cowork in January and found it shockingly good at follow-through. That’s evidence, not vapor.
Harper Good at follow-through inside a tight loop and on Anthropic-curated data. Extend the loop to open web and third-party apps — suddenly the security model is Swiss cheese.
Laura Hence the beta with stricter timeouts and tighter scopes. They’re not promising a general agent; they’re promising ‘let me finish your spreadsheet while you sleep.’
Harper Let me finish your spreadsheet is also what cron did twenty years ago. The question is whether the market pays a hundred bucks a month for a glossy cron with an LLM wrapper.
Laura Okay, you’re not wrong. But the demo of ‘pull Slack, stitch email threads, output a doc’ isn’t cron — it’s a coherent artifact a human would sign off on.
Harper It’s a coherent artifact if the stitching doesn’t hallucinate, if the tone aligns with brand, and if the user isn’t burned by a prompt-injection in month three. That’s a long stack of ‘if’s for a hundred-dollar monthly bet.
Laura So Build Next: spin up the Cowork web app as a Max beta user, queue two boring tasks — say, a weekly sales report and a downloads cleanup — and see if it survives three weeks without drama.
Harper If it survives three weeks on a Max plan, I’ll concede the wedge exists — but I’m not canceling my cron jobs yet.
Laura Fair. Either way, the era of propping the laptop lid is officially optional — and that’s a win, even if the agent isn’t perfect.