Anthropic ConwayThe rumors have pushed the Always-on Agent track a step forward. If the leaked information is true, this project will not just add an extra entry to Claude, but will consolidate the browser, connector, Claude Code and external triggers into a resident worker thread.
What does the rumors of Conway really point to?
What the market is most interested in Conway is not a new name, but a combination of "separate UI + browser operations + connector + webhook calls". Most AI products today are still at the conversational assistant stage. When you ask a question, it responds; when you leave, it stops. Always-on Agent means the opposite: tasks can be suspended, called back, continued execution, and even relayed among different tools.
This is why the spreading extension mechanism and CNW ZIP are more worthy of attention. If Anthropic really designs a unified package for custom tools, UI tabs, and context handlers, Claude's role will change from a model in a chat box to an installable, triggered, and orchestrable agent runtime. For developers, the focus is not just "whether the model can be invoked", but "whether the business process can be stuffed into Claude's work loop."
What Anthropic lacks is not the model but the entrance
From the perspective of product path, Anthropic's actions in the past two years have not been scattered.Claude Codeis responsible for development tasks, the connector is responsible for bringing in enterprise data and SaaS, and the browser capabilities and Cowork/Dispatch are testing whether Claude can continue to do things after people leave. If Conway is implemented, it will essentially build a unified scheduling interface on top of these already rolled out capabilities.
This step is critical. The AI industry is moving from a "stronger dialogue model" to a "longer chain of task execution." Whoever can keep the agent online, understand the context, and complete tasks across tools will be closer to the next main entrance. If Anthropic only protects models and APIs, it will be difficult to occupy user desktops for a long time; it needs a resident product that pushes Claude from the "best model" to the "first workbench to be opened."
Always-on Agent will rewrite first
The first thing to be rewritten is not ordinary chat, but two types of high-frequency work: one is development and operation and maintenance, and tasks can be naturally disassembled, callbacks, and triggered by webhooks; The other type is knowledge work, such as inspecting reports, organizing meeting materials, following up on sales leads, and moving information across systems, which inherently relies on browsers, connectors, and repetitive processes.Always-on AgentWhat really eats is the workflows that are always clicked on repeatedly by humans.
There will be no fewer questions. Once a resident agent has browser control, external connections, and custom extensions, the security boundary is no longer just about "whether the model will talk nonsense", but about "which systems it can move, who will audit, and how to roll back if something goes wrong." Anthropic has always regarded security as the core of its brand. If Conway is really pushed to the forefront, rights management, approval links and logging capabilities will be more important than any fancy interface design.
Conway is now more like a product puzzle exposed through the source code, but the direction it reveals is very clear: the next stage of AI competition is not just who answers better, but who can continue to do the work while you are not staring at the screen. ForAnthropic, after Claude Code, the real battle may have just begun. At present, public information is still limited, and follow-up progress deserves continued attention.
