The agent left the laptop. The labour question stayed behind.

Anthropic put Claude Code on the web and the iPhone this week. The headline is access: an autonomous coding agent that used to require a terminal now runs from a browser tab or a phone screen. The framing in the trade press is that this changes who can write software. I think it changes something else first, and the access story is hiding it.

What actually shifted is where the work happens, when it happens, and who is watching it while it happens. The same week, OpenAI's Codex gained the ability to operate a locked Mac, meaning the agent runs while the user is asleep, away, or simply not paying attention.¹ Put the two together and the product category that emerged in May 2026 is not "coding on your phone." It is coding without you in the room. That is a different thing, and it has different consequences.

Start with who this is being sold to. Tom's Guide, which BuildFastWithAI cites in its summary of the launch, says the web and iPhone version is "about to change how you vibe code."² Vibe coding is a term popularised by Andrej Karpathy in early 2026 for a casual, non-professional style of software work: describe what you want, accept what the model produces, ship it. The audience for that framing is not the employed senior developer at a bank. It is the hobbyist, the founder, the student, the operator who wants software but does not write it for a living.

This is the segment Anthropic is pricing for. In the same product cycle, Anthropic has been testing a $100 per month Claude tier alongside its existing $20 Pro tier.² A hundred dollars a month is not a professional engineering procurement decision; an engineering team buys seats by the dozen and negotiates. It is a disposable-income decision for an individual user who wants more agent than the entry tier provides. The mobile surface and the higher price point are aimed at the same person: someone with money, some technical confidence, and an appetite to direct software work without doing it themselves.

Now ask who is exposed. If the buyer is the founder with $100 a month, the worker affected is the junior developer, the contract coder, the offshore engineering team whose job was to turn that founder's specifications into working code. That labour relationship is the one the agent substitutes into, not the senior staff engineer's job. The senior engineer still has the multi-monitor desk, the full IDE, the code review authority, the production access. The iPhone interface is not coming for them this quarter. It is changing the economics of the rung below them.

This is the standard pattern of consumer-AI rollout, and it is worth naming plainly. The capability lands in the hands of the people with the most disposable income and the least exposure to its labour consequences. The people whose work it most directly substitutes for are neither the target market nor, typically, the readers of the announcement. They find out later, through their managers, through reduced contract volume, through job postings that no longer appear.

The "off the laptop" framing also changes the supervision question, and that matters more than the access question. When a coding agent runs in a terminal on a developer's machine, a human is present. They see the diff, they catch the bad suggestion, they intervene before the agent commits a change that breaks production. That human-in-the-loop assumption is doing a lot of unspoken work in how we talk about these tools as "augmentation."

A web-hosted agent operating on a sandboxed copy of a repository, accessed from a phone, with the user reviewing output asynchronously, is a different supervision regime. A locked-Mac agent running overnight is a different regime again. Anthropic's web version runs inside its managed sandbox with restricted file and network access,³ which is a sensible engineering choice and also a tell: the company is designing for a world where the agent operates with meaningful autonomy, and the controls are technical guardrails rather than continuous human judgement.

If the junior developer's job was to supervise and implement what the senior engineer specified, what is that job when the agent runs on a phone the senior engineer checks between meetings, or on a locked Mac while everyone is asleep? The honest answer is that we do not yet know, and the honest follow-up is that the people who will find out first are not being consulted about it.

The contrary case is worth taking seriously, and it does not rescue the framing. GitHub Copilot, Cursor and Amazon CodeWhisperer have been browser- and IDE-accessible for months or years; the platform surface itself is not novel.² Professional engineering teams will continue to work from desks with full toolchains, version control integrations, and team-level review flows. The iPhone is an initiation and monitoring surface, not a production environment, and the gap between "start a task from your phone" and "developers replaced by phone-accessible agents" is large.

All of that is true, and none of it addresses the actual shift. The shift is not that production engineering has moved to mobile. The shift is that agent runtime is decoupling from human attention, and the pricing model, per-seat, per-month, is being probed at exactly the moment the underlying economics stop matching it. When one seat can run persistent tasks continuously across web, mobile and locked desktop, the seat is no longer the unit of value. Consumption is. Whoever absorbs that cost shift — the buyer, the platform, or the displaced worker through reduced demand for their time — is the distributional question of the next year of this category.

What to watch. Whether the $100 tier holds, gets raised, or gets replaced by usage-based pricing. Whether enterprise policy permits locked-machine agents in the environments where professional developers actually work, or pushes them back into sandboxes. Whether the hobbyist and founder cohort the mobile surface targets generates enough revenue to subsidise the heavier compute these agents consume, or whether the price floor moves up. And whether anyone, regulator, union, professional body, asks the people whose work is being restructured what they think before the restructuring is finished.

The agent leaving the laptop is the product story. The labour question stayed behind, and it will be answered whether or not the people most affected get a voice in the answering.

Footnotes

¶