How Anthropic is turning Claude into an ‘always-on’ agent — and what it learned from OpenClaw

Anthropic is turning up the dial on its efforts to make Claude a system that can carry out tasks without constant input, introducing a trio of new features this past week that let the model run in the background, act across a user’s machine, and be controlled remotely.

The direction can be seen in the context of OpenClaw, the viral open-source agent that showed how language models could operate a user’s computer directly, navigating applications and carrying out tasks across systems.

That work has already drawn attention from Anthropic’s rival OpenAI, which hired OpenClaw’s creator in February to work on what he described as “bringing agents to everyone.”

Claude: Ready for Dispatch

Last week, Felix Rieseberg, a member of technical staff at Anthropic, introduced Dispatch as part of Claude Cowork, a recently-released agent mode in the Claude desktop app that allows users to assign tasks and return later to completed work.

Within Cowork, Claude executes tasks on the user’s machine, working across files, the browser, and connected tools with user approval. Dispatch builds on that by allowing a single session to persist, so work can continue in the background while the user is away.

“One persistent conversation with Claude that runs on your computer,” Rieseberg wrote. “Message it from your phone. Come back to finished work.”

For example, a user might leave a task running on their desktop, then open their phone and send a message into the same session—asking Claude to check on progress, adjust the task, or kick off a follow-up while it continues working in the background.

The feature is rolling out as part of Cowork’s research preview for both Max and Pro subscribers. It requires the Claude desktop app paired with a mobile device, and the host machine must remain running for tasks to continue.

It’s worth noting here that while OpenClaw showed what’s possible when a model is given broad access to a user’s computer, it also exposed a range of security concerns, from prompt injection attacks to credential leakage and vulnerabilities that allowed external control of the agent.

Anthropic’s incarnation, meanwhile, introduces more structure around that interaction. Within Cowork, tasks run in a sandboxed environment, with Claude requesting approval before accessing files or taking actions.

This theme has come up repeatedly in online discussions, where the community has focused on how access is controlled once these systems are given real responsibilities. Connor Grennan, CEO of AI Mindset, pointed to the appeal of getting an always-on agent that can run tasks in the background without the risk associated with OpenClaw.

“It’s the ‘always-on’ agent vibe from OpenClaw without the technical setup or security nightmares,” Grennan wrote on LinkedIn.

Others agreed, suggesting that the main issue at play here was that of trust.

“Dispatch isn't competing with OpenClaw on features – it's competing on trust,” they wrote. “Anthropic is betting that most people won't hand their entire machine to an open-source agent they can't audit, especially once real work depends on it.”

Channel surfing with Claude

Separately, Anthropic also introduced Claude Code channels, which extend how users can interact with a running session beyond the native app. Dispatch is positioned as a general productivity feature for assigning and tracking tasks, whereas channels are aimed more directly at developers, offering a programmable layer for interacting with a live Claude Code session. Channels open that same session up to external services such as Telegram and Discord.

Messages from those platforms — including CI results, alerts, or webhook events — are forwarded into the live session, with Claude responding in the same thread. This allows developers to route system events directly into Claude, check on progress, or send instructions without returning to the interface. Built on Anthropic’s Model Context Protocol (MCP) system, channels support multiple entry points into the same session, including custom integrations that can pass information or trigger actions while Claude is running.

For example, a developer might have Claude running a test suite in the terminal, then send a message from Discord or Telegram that appears directly in the session, allowing them to check on the run or intervene while the tests are still executing.

Claude learns to use a mouse and keyboard

In what might prove to be a particularly useful upgrade, Anthropic also introduced a feature that allows Claude to operate a user’s computer directly, extending what it can do once a task has been assigned.

Claude can control the mouse and keyboard, open applications, navigate the browser, and interact with software in much the same way a user would, allowing it to work across tools that don’t have direct integrations. Instead of relying on APIs, it can carry out tasks through the interface itself, moving between applications as needed.

As with Dispatch, access is gated. Claude only gets access to the folders and connectors a user explicitly permits, and for computer use it asks permission before accessing each application. Anthropic also says Claude shows its plan and waits for approval before taking significant actions.

For example, a user might send a message asking Claude to export a pitch deck as a PDF and attach it to a calendar invite, with Claude opening the relevant app, generating the file, and completing the task while they are away.

Notably, computer use will work with both Claude Cowork and Claude Code on desktop, meaning it has a wide range of use-cases for developers too.

“You can ask Claude to click all the buttons in a legacy app that you'd like to automate – or use it to help debug a native app you're working on,” Rieseberg noted.

Anthropic is also positioning the feature as deliberately cautious in how it operates, reflecting how early this category still is.

“The entire computer-use field is early – Claude will move slowly and deliberately, much slower than a human does,” Rieseberg said.

OpenClaw, in many ways, demonstrated the power of giving a model broad control over a machine. But it also exposed how difficult that becomes to manage once the model is interacting with real systems. Anthropic’s version keeps the basic idea intact, while wrapping it in permissions, checkpoints, and defined entry points – a version designed for real machines, with real files, under real constraints.