New Intelligence Report
Editor: KingHZ
【New Intelligence Insight】Anthropic's "always-on" resident agent and brand-new UI exposed! Claude will take over your computer background 24/7 to execute tasks autonomously, truly freeing your hands.
At three in the morning, Silicon Valley was in an uproar.
In just a few hours, a post from last month had the entire tech world losing sleep.
Hidden within an update to Anthropic's npm package was a 59.8 MB source map file.
What does that mean?
It means exactly this: 512,000 lines of unobfuscated TypeScript source code, 1,900 files, practically the entire internal implementation, were left completely exposed on the public web.
The reason was so absurd it sounds like a joke: a misconfigured line in .npmignore.
But what really got everyone on edge wasn't the "code leak" itself.
It was a codename that jumped out from the leaked code, one never before mentioned in any public documentation: Conway.
That single word changed everything instantly.
Because digging deeper, what people found was a new species taking shape—
Claude is transforming from a chat window into a 24/7 background-resident, continuously running agent ready to take over tasks at any moment.
You close your browser, it keeps working. You lock your phone, it doesn't stop. You go to sleep, it continues running tasks in the background.
Triggered by an external event, it wakes itself up, executes tasks on its own, and sends you a notification when it's done.
Reading this, you should realize:
1. What Anthropic wants to build is not just a smarter chatbot.
It wants to become the entry point for the next-generation operating system.
This is Not an 'Upgrade', This is a Species Change
For the past two years, the entire large model industry has been trapped inside a chat window.
You type a message, it replies.
You ask a follow-up, it supplements its answer.
No matter how powerful, it's still just "you ask, it answers."
Close the window, the task pauses; lose the context, the intelligence resets to zero.
The AI, however brilliant, is just a clicked-open browser tab.
But Conway doesn't follow that logic.
From the leaked code structure, it is an independently running Claude instance environment.
It has its own sidebar UI, divided into three core areas: Search, Chat, and System;
It supports webhook wake-up;
It can run Claude Code;
It can control Chrome;
It can send system notifications;
Behind it lies a background task queue, persistent session state, and a system-level process daemon.
Translated into plain English:
Previously, using AI was like "hiring a temporary consultant." Now, what Anthropic wants to give you is a "digital employee who never clocks out."
You give an instruction during the day: "Keep an eye on this GitHub repo for me, and pre-review any new PRs."
Then, you shut down your computer, leave work, and go to bed. The next morning, the review results are already written.
You don't need to call on it; it's been waiting in the background, ready to act on its own at the right moment.
One is like a remote control. The other is like a butler. The difference is so vast it doesn't belong to the same era.
Anthropic's True Ambition
AI Living Forever Inside Your Device
Many people's first reaction might be: Isn't this just a slightly more powerful AI assistant? Isn't this just "Claude's version of OpenClaw"?
No. Absolutely not.
The truly terrifying part is that what Conway exposes isn't a single-point capability, but platform-level capability.
Anthropic has also hidden an extension system in this release.
Anthropic has pre-built an Extensions area, supporting the installation of custom tools, UI tabs, and context processors.
Extension packages even have a new format: .cnw.zip.
The signal is crystal clear—Anthropic is building an ecosystem.
Frankly, Anthropic doesn't just want to make Claude "a bit more useful."
It wants to turn Claude into a resident agent platform that can be installed, woken up, extended, and managed over the long term.
Look at another detail: the code contains parts that are commented out but not deleted, hinting at future support for cross-device state synchronization.
A task you assign on your phone can be continued by your computer. A thread you start on your computer can be persistently followed from your mobile device.
A task is no longer bound to a specific window, a specific device, or a specific session.
Once this layer is connected, what Anthropic captures isn't just the entry point to an AI assistant.
It's an "agentic operating system layer" that runs through all your devices, all your workflows, all your digital behavior.
That is its true scheme.
Conversational AI Has Hit its Ceiling
Many haven't yet realized it, but this competition has already changed lanes.
What was the past race about? Parameters. Long context. Who sounds more human. Who answers more elegantly.
What will the next race be about?
Who can actually get things done for you.
Because users are already getting fatigued with "chatting."
Of course, you'll still ask questions. But what you truly want has never been an "AI that speaks more nicely." What you really want is:
I don't have to watch it. I don't have to push it. I don't have to repeatedly explain myself. I don't have to write prompts over and over.
Ideally, I state my goal, it runs in the background on its own, and finally gives me the result.
This is the second half of the AI game.
Therefore, what Conway truly breaks isn't just a product form. It breaks the AI interaction method we have all accepted by default until now.
The chat dialog is the starting point for AI, but it absolutely will not be the finish line.
Anthropic's Move Here is Ruthlessly Bold
Looking at Conway within the context of Anthropic's moves over recent months, it's actually a very clear roadmap.
End of 2025: Claude Code is released. First, get developers accustomed to a "locally, continuously running AI programming assistant."
Subsequently: Claude Cowork goes online. Persistent agent workflows begin to take shape.
March 17th: Claude Dispatch opens to Max users. Persistent threads start running on mobile and desktop, and even if you close the app, the tasks continue.
April 8th: Managed Agents enter public beta. Sandbox execution, checkpoints, credential management, permission control, and end-to-end tracing—an entire suite of infrastructure is completed.
See that? It's a roadmap with a clear objective:
Chat dialog → Background process → Resident platform → Operating system entry point.
Step by step, Anthropic is pushing Claude from "being able to answer questions" towards "doing the work for people."
Those familiar with tech history probably already have goosebumps reading this.
In 1995, Microsoft bundled Internet Explorer into Windows.
It wasn't about making a better browser; it was about redefining what a user first encounters after turning on a computer.
Thirty years later, whoever binds AI into the operating system first is most likely to secure a ticket to the era of agents.
Clearly, Anthropic has already entered the arena.
The Biggest Dark Humor
The Most Sensitive Roadmap, Ruined by a Single Line of Configuration
The most absurd part of the whole thing is this:
In terms of AI safety, Anthropic invests extremely heavily.
They've published many papers, and they make the most frequent statements. Yet, their most sensitive product roadmap was ultimately ruined by the most ordinary packaging oversight.
Not a top-tier hacker. Not a sophisticated attack. Not a zero-day exploit. Just one configuration option.
It's almost like a slap in the face to the entire industry:
You can train your models to be stronger, make your permission system more granular, and讲述 a more complete safety narrative—and still lose to the simplest human error in the end.
And what's even more alarming is that once these kinds of resident agents truly get into the base layer of a device, their attack surface will swell dramatically.
Before, humans invoked AI via APIs; in the future, AI might invoke the world via webhooks.
Before, we worried about models talking nonsense. In the future, we'll have to worry: will it actually do something in your name while you're not looking?
This is no longer a change that can be summarized as a "chat software upgrade."
This is the dusk of the chat dialog, the dawn of the agent. This is a redistribution of computing entry points, task execution authority, and digital sovereignty.
AI is no longer content to stay in the chat window.
It's starting to want to live in your device, take over your tasks, and rewrite your workflow.
The real question has finally arrived before us:
When a 24/7 online digital counterpart emerges—one that knows your computer better than you and understands your workflow more intimately—
Are you ready to hand over your keyboard?
Reference:
https://www.testingcatalog.com/anthropics-works-on-its-always-on-agent-with-new-ui-extensions/