Manus agents get their own wallet and phone number in a personal app

Software used to be a tool that people operate. Agents change that, because they act on their own. So the buying question shifts. Ask what the agent is allowed to touch, not only what it can do.

What Manus released

Manus has rebuilt its AI agent platform. InfoWorld reports that the company calls Manus 2.0 "a new architecture, new products, and new capabilities." It runs on web, desktop and mobile.

There is some history here. Meta said in December 2025 that it wanted to buy Manus. China's National Development and Reform Commission stopped the deal in April. By August, Manus said it was independent again. Now the two compete. Meta has its Muse assistant and a new Enterprise Platform.

Two parts of the release speak to different readers. Cue is a standalone app for personal agents. Manus says that in Cue, "each agent has its own email, phone number, wallet, and computer." Those agents can spend within set limits. Several can share a group chat and pass work to each other.

How the machinery works

Cascade is the harness. That is the layer that decides when an agent is called and how several agents coordinate. Manus says it keeps projects light and adds specialised skills only when needed. That matters because AI services bill by tokens, the small units of text a model reads and writes.

These figures come from Manus. They compare Cascade with the company's own earlier system. Manus did not disclose the setup or the tasks. Treat them as a claim to test, not a benchmark.

Gartner's Anushree Verma says the orchestration layer is becoming a bigger differentiator for enterprise deployment. She adds that it complements the underlying model and does not replace it.

Automations are the second change. Before, work ran on a schedule. Now an agent can also start when a connected service reports something. Examples include incoming email, a Slack message, a move in ad metrics or a calendar item.

Who carries the risk

Picture what these features could add up to. An agent wakes on an incoming message, holds credentials and runs overnight. That looks more like a new hire than a tool. It is like giving a contractor a badge, a company card and a laptop on day one.

This is an illustration, not a description of how Manus sells the product. The source does not say these features combine into one enterprise agent. But the picture shows why governance matters.

Verma was direct: "The capability is becoming enterprise-relevant, but the autonomy is ahead of the governance." She advised treating such systems as "untrusted or semi-trusted automation with strong isolation, approval gates, detailed telemetry, and narrowly defined authority."

Here AI acts as a risk multiplier. Email and chat are the triggers Manus names. They are also channels that outsiders can write to. The source reports no attack on Manus. Still, an agent that acts on incoming messages takes on the trust questions those messages bring.

Questions to put to your team

Verma and Grover offer a workable checklist. Turn it into questions before any pilot starts.

First, where are prompts, task state, logs, backups and connector credentials processed and stored? Include model providers and other subprocessors. Second, can a workflow be rebuilt outside the platform? Has anyone tested that?

Third, for higher-impact uses, are credentials narrow and short-lived? Is there audit logging outside the vendor's platform? Do consequential actions need a human approval? Can access be revoked quickly?

Fourth, ask what any cost claim rests on. Request the setup and tasks behind a savings figure. Manus has not shared them.

Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: InfoWorld.