If you asked an AI practitioner last year "when can an agent work through the night for me," the most honest answer was probably "you'll need to leave a computer running."
That answer is obsolete. On July 7, Anthropic announced Claude Cowork's formal launch on mobile and web. Desktop tasks can now be checked and taken over from a phone at any time. Chat and Cowork merge into a single entry point, with usage doubling rolling forward. It sounds like a routine product update. But Anthropic engineer Felix Rieseberg's framing deserves a close read — he called it Cowork assembling all three critical puzzle pieces for the first time: precise context acquisition, advanced loops for long-horizon tasks, and complete freedom from device constraints. The first two were each individual bottlenecks; this time, they are welded together. This is not a feature update. It is a paradigm shift from "desktop tool" to "cloud-native service."
The Three Puzzle Pieces
1. Precise context acquisition
An agent's first problem was never "can it do the task" — it is "does it know what you're talking about." Tell Cowork "organize the Acme project materials into a prep doc, draft the follow-up email but don't send it, make a retro outline for the meeting," and it cannot operate on literal instructions alone. It needs to know: which project is Acme? Where are the materials scattered (which Slack channel? which emails? which calendar meetings?)? What template does "prep doc" use? Who receives the "follow-up email"?
The old solution was prompt engineering — writing an enormous instruction, pouring all context in. Users were capable but unwilling: re-typing background for every task turned the agent into another tool needing maintenance rather than a colleague who does the work. Cowork's new approach: let the agent "probe" available connectors itself — Slack, email, calendar, connecting one by one — then search chat history, filter emails, and assemble context autonomously. You stop supplying background; it finds it.
2. Advanced loops for long-horizon tasks
The agent's second problem: tasks beyond fifteen minutes start "drifting." It forgets the initial goal halfway through, or grinds on a branch without reporting main progress. Loop engineering became the hottest agent-direction topic of H1 2026 precisely because the industry discovered that enabling long-duration autonomous work is not an endurance problem — it is an architecture problem. Cowork implements a complete work circuit: receive brief → plan autonomously → find tools → pull data → produce documents → pause at decisions for human sign-off → leave output as draft. The key is the "pause at decisions" step. It does not grind to the end and dump results on you — it stops at judgment nodes and pushes the question to your phone: a notification pops, you approve from the subway, it continues. This is not simply "splitting tasks into steps" — it establishes a human-machine collaborative rhythm: the agent runs the tedious parts, the human handles the judgment calls. Each to their own.
3. Device decoupling, cloud takeover
This is the update's most fundamental change. Previously, assigning work to AI from a phone required the computer to stay on as the host. Network down, work stops. Now tasks run on Anthropic's cloud — whether you are present or your computer is on does not affect completion. Ramp customer-success manager Armmand Hosseini's experience is representative: he was building a client-tracking dashboard with Cowork, had to leave for a flight mid-task, and when he landed, standing at baggage claim, he pulled out his phone — the cloud session was still running. Cowork popped a decision awaiting his sign-off; he tapped confirm, closed the phone, picked up his bag, and walked. The value of this scenario is not "the phone can control the computer." It is "the computer no longer needs to be online." The screen is just the interface you use to interact with the agent — no longer the home where it runs.
From Sparring Partner to Free Range: Three Stages of Agent Evolution
Over the past two years, agent deployment has crossed three stages, and Claude Cowork marks the third. Stage one: sparring partner (Copilot mode, 2024-2025). The human leads; AI assists — code autocompletion, smart document suggestions. AI is a trained assistant: you decide, it accelerates execution. The bottleneck is obvious — it cannot act independently. You stop, it stops. Stage two: shift replacement (agent mode, 2025 - H1 2026). Humans can delegate tasks; agents execute independently. OpenClaw's early-year surge captured this moment — say one sentence, it opens software on another computer, processes files, runs scripts. But the essence is "remote desktop plus an AI brain": the agent runs on your computer, with your permissions, using your compute. Close the lid, it goes offline. The bottleneck: agent and device are bound at the hip. Stage three: free range (Cowork mode, mid-2026-). Humans delegate, AI decomposes, AI finds its own tools, AI verifies its environment, AI runs on cloud, humans supervise from a phone. The key change is not that models got stronger — it is that the execution environment moved from desktop to cloud-native architecture. The agent went from "a program on your computer" to "a service on the cloud." You watch the World Cup, catch flights, sleep — the work runs. The phone becomes the command deck; the desktop remains the deep-work home base but is no longer indispensable.
The Endgame: Ubiquity, Not Intelligence
Understanding this three-stage evolution makes the other giants' moves far clearer. OpenAI Codex weekly actives passed 5 million, with knowledge workers at roughly 20% and growing three times faster than developers. The key signal is not "many people use Codex" — it is that "knowledge workers are becoming agents' largest incremental market." People who write code are finite; people who do analysis, reports, and management are infinite. Microsoft's analysis of 100,000 Copilot conversations points the same way: 49% of usage supports analysis, evaluation, and problem-solving — cognitive work unrelated to coding. Microsoft's recent reorganization — merging consumer and enterprise Copilot into a unified app, launching the AutoPilot agent — is essentially the same bet on "ubiquitous agent services," not "one more button in Excel." Google's Gemini Agent Mode is contesting the same territory. Three giants, three data points, one conclusion: the agent's true incremental market is not in the code editor — it is in ordinary workers' daily busywork.
Anthropic's moves are the most aggressive. They saw the market and bet on the path to it: freeing the agent from the device. January: desktop launch. February: expansion to general office workers. Then Dispatch enabling remote task delegation from phones. Now: mobile and web, scheduled tasks free of devices. From desktop to palm to cloud — a flag planted at every interface where work happens. The logic deserves careful thought from every agent product team: model capability will converge. GPT-5, Opus 4.6, Gemini 3.0 — the gaps are narrowing. You cannot build a durable advantage on "my model is 5% smarter." But infrastructure gaps will not converge. If one agent keeps running for 6 hours after you close your laptop and pushes decisions to your phone, while a competitor's product requires you to stay at the screen — which do you choose? The endgame of this competition is not whose agent is smarter. It is who becomes part of ordinary people's daily routine first.
What To Do
If you are a product manager: stop asking "can our agent do X" and start asking "can our agent continue doing X after the user leaves the device." Cloud-native, offline-capable, mobile-commanded — these are 2026 agent-product baseline configurations, not bonus features. If you are a developer: focus on loop architecture and agent-runtime design. Cowork's key innovation is not the model — it is the task circuit, context probing, and decision-interruption mechanisms. These engineering practices transfer more easily than model weights and differentiate experience more sharply. If you are a knowledge worker: you do not need to learn prompt engineering. You need to learn: throw the task out, set completion criteria, check progress on your phone, appear only when decisions are needed. Products like Cowork are turning "using AI" from a skill into a habit. If you are a founder: the agent infrastructure layer has enormous whitespace. Device-agnostic agent runtimes, cross-service context management, human-AI decision routing — these are the "more important than models" engineering questions, and the giants have only just started defining the answers.
References: Anthropic Blog — Claude Cowork on web and mobile · Synced, "Close your computer; it works through the night" · 36Kr, "Claude Cowork major update today"