An Agent With Its Own Phone Number and Wallet: Manus 2.0 Rebuilds Itself Around Cascade, Cloud Computers, and Cue

On the morning that San Francisco braces for OpenAI’s DevDay, the most interesting agent launch of the week came from the other side of the Pacific. Manus — the Singapore-based startup behind one of 2025’s most hyped AI agents — shipped Manus 2.0, and it is not a version bump. The company calls it “a new architecture, new products, and new capabilities,” and for once the marketing language undersells what arrived.

The release bundles five things: Cascade, a rebuilt agent harness that the company says used 23.2% fewer tokens, finished tasks 28.2% faster, and cost 32% less to run than its previous system; Cloud Computer, a dedicated persistent environment for projects that outgrow a single machine; Automations, which evolve scheduled tasks into event-triggered workflows; Manus Studio, a desktop-app-turned-shared-workspace with an honest-to-goodness video editor and game development environment; and Cue, a standalone personal-agent app whose agents each carry their own email address, phone number, wallet, and computer.

That last item is the one worth paying attention to.

The architecture: agents that travel light

The core engineering story of Manus 2.0 is Cascade. The previous Manus architecture, like most agent frameworks, tended to load a broad set of capabilities into every task — an approach that makes simple work slower and more expensive than it needs to be. Cascade inverts this: a project starts light, and specialized capabilities are brought in only when the work demands them.

“The brief, the page, the video, and the automation stay part of one connected project,” the company writes — meaning context is shared across artifacts rather than fragmented across sessions. In one tested configuration, the company reports, Cascade needed 23.2% fewer tokens and 28.2% less task time to do the same work, translating to a 32% cut in operating cost.

Those numbers matter beyond Manus’s own margins. Token efficiency has become the defining economics of the agent era: as agents run longer and touch more tools, the cost of wasted context compounds. Anthropic-style managed agents, OpenAI’s anticipated DevDay announcements, and now Manus are all converging on the same insight — the winning agent platforms will be the ones that stop paying frontier-model prices for boilerplate steps. A 32% cost reduction claimed on real workloads is a serious swing in a market where agent bills are the new cloud bills.

Cloud Computer and Automations: agents need a place to live

Two of the new primitives address an awkward truth about today’s agents: they are homeless. Close your laptop and the work stops.

Cloud Computer gives a project “a place of its own” — a dedicated cloud environment for the things that must stay up: a multiplayer game server that survives a closed lid, or “a permanent home for an automation” that runs around the clock. It is the infrastructure half of a bet that agents will increasingly need persistent, always-on compute rather than ephemeral sessions.

Automations extends the same idea to time and events. Scheduled Tasks, the old mechanism, could only start work at a set time. Now “a new email, a change in ad performance, a calendar event, a Slack message, or a Notion update can all trigger a task.” The setup interface is strikingly simple: tell Manus in a single prompt what to watch and what to do, and it configures and runs the workflow itself. This is the trigger-action programming model of IFTTT and Zapier, except the “programming” is one sentence.

Studio: from finishing tasks to making things

The second half of the release is aimed at output quality. Manus Studio upgrades the old desktop app into what the company calls “a shared workspace for people and AI,” with professional tooling for documents, spreadsheets, PDFs, slides, websites, code, games, and video.

The Video Editor is the standout. Instead of the regenerate-from-scratch loop that defines most AI video workflows — want a warmer song? Re-prompt and pray — Manus exposes the timeline. The clips, images, text, motion graphics, and audio it used are separate editable pieces: swap the song, drop in your own footage, move an animation, export when happy. You can also hand the timeline back and let Manus take another pass from your edits. An “Alchemy mode” goes further, letting AI act as creative director with video and code generation working in tandem, with the company claiming noticeably better pacing and visual control.

Game Dev applies the same philosophy to games: build from a playable template, iterate in Max Ultra mode, hand-edit anything (“if one villager’s hair should be silver instead of brown, select that villager and change it”), then publish as a playable website. The multiplayer problem — the hosting, networking, and upkeep that stall most hobby game projects — is solved with a few clicks by attaching a Cloud Computer as the always-on server.

Cue: the agent gets a phone number

Then there is Cue, the piece that gestures at where all of this is heading.

Cue is a standalone app for personal agents, on phone and desktop, built on Manus’s infrastructure. “Give it a cue, and it gets to work.” What distinguishes it from a chatbot with memory is the identity layer: every agent gets its own email, phone number, wallet, and computer. It can send messages, pay within a budget you set, and complete tasks on its own machine. It can take your calls and leave a summary. Multiple agents can be placed in a group chat with a shared goal, handing work to one another — one researches venues, another shortlists, a third drafts the deck — while you set direction and make the final call.

There is even a physical-world hook: at a restaurant, scan the QR code and your agent can order for you or hold your place in line.

It is easy to see the seam between vision and product here — Cue is in early access, free with an invite code (MEETCUE, limited and first-come-first-served), and services like restaurant ordering depend on partner integrations that are thin at launch. But the design commitment is telling. Giving an agent its own phone number and wallet is a claim that agents are becoming economic actors, not just tools — counterparties that can be reached, that can transact, and that need their own accounts to do so responsibly. It also concentrates a remarkable amount of trust in one platform: an agent with a budget, a phone, and your credit line is a rich target for exactly the kinds of agent-security incidents that have dominated this month’s headlines.

Context: from viral invite codes to a platform bet

Manus has come a long way from the invite-code frenzy of spring 2025, when codes resold for over $1,000 and the company’s parent, Butterfly Effect, quintupled its valuation to nearly $500 million. It has since relocated to Singapore, and in mid-September it was reported to be raising $500 million at a $4 billion valuation. Manus 2.0 is the product evidence for that price tag: not a demo of a smarter model, but an attempt to own the full stack of the agent era — the harness, the compute, the tools, and the identity layer.

The timing, hours before OpenAI’s DevDay in San Francisco, is clearly deliberate. OpenAI is widely expected to unveil its own “managed agents” today. The two companies are making adjacent bets: OpenAI from the model side, Manus from the product side. Whichever architecture wins, the direction is the same one Manus named in its launch post — a world where the interesting question is no longer “can AI do this task?” but “what does AI need to own in order to finish it?”

An email address, a phone number, and a budget, apparently.

---

Manus 2.0 is available now, with the Cue iOS app “coming soon after App Store review.” Early access to Cue is free with the invite code MEETCUE.