Manus 2.0 Gives Each Agent an Identity Layer

Manus 2.0’s most consequential move isn’t a model upgrade — it’s giving each agent an email, a phone number, a wallet, and a computer, and everything that creates downstream.

The only source for this piece is the Manus company blog, which is a promotional document. The efficiency percentages come from what the company calls “one tested configuration”, measured by Manus against its own previous system. They are not a benchmark, not independent, and not claimed to be typical, and no workload, model or run count is disclosed. The agent wallet pays within a budget the user sets, so the spending is delegated and capped rather than autonomous. Cue is in early access and invite-only, with iOS still pending App Store review. Nothing here is investment advice.

What Happened

On Monday, September 28, Manus published a company blog post announcing Manus 2.0, a new standalone app called Cue, and an upgraded orchestration layer it calls Cascade. The source for everything that follows is that promotional document alone — no independent benchmark, no user count, no revenue or pricing figure appears in it, and none is supplied here.

The sentence in the announcement worth stopping on is about identity: “each agent gets its own identity” — specifically its own “email, phone number, wallet, and computer, so it can send messages, pay within the budget you set, and see a task through on its own machine.” The qualifier in that sentence is load-bearing and stays attached throughout: the wallet pays within a budget the user sets. That is delegated, capped spending — not autonomous or unlimited spending by the agent.

On availability: Cue is in early access, invite-only, open to a limited number of early users on a first-come, first-served basis with the code MEETCUE. iOS is still pending App Store review. This is not general availability. Manus Studio, also announced, adds documents, spreadsheets, PDFs, slides, websites, code, games, and video as output types, with two new professional environments — Video Editor and Game Dev.

The Structural Read

The Map of AI places companies across nine layers — from infrastructure and model providers down through orchestration, applications, and the account and identity rails that sit between AI output and the real world. Manus 2.0 is a deliberate move down that stack.

What Manus is describing is structurally identical to employee onboarding. When an organisation hires someone, the first things it issues are an email address, a phone number, a corporate card with a spend limit, and a machine. Those four things are not perks — they are the minimum credentials that let every other system (banks, vendors, communication platforms, internal tools) recognise the new hire as a valid counterparty. Manus is issuing the same four primitives to each agent.

That move does something specific to where the hard problems live. The capability question — can the model reason well enough to complete the task — gets easier to answer with time and compute. The account question — who authorised this spend, who carries the liability if the payment fails or the message misleads the recipient, what the counterparty on the other end of the phone number believes it is dealing with — does not get easier automatically.

It gets harder as agents proliferate. The blog addresses none of those questions, and this piece answers none of them either. But the design creates them.

The Cascade efficiency numbers reinforce the orchestration ambition, but they require careful reading. In one tested configuration, measured by Manus against its own previous system — not against a third-party benchmark, not claimed to be typical — Cascade used twenty-three point two percent fewer tokens, finished tasks in twenty-eight point two percent less time, and cost thirty-two percent less to run. Indexing the previous system to one hundred, this publication’s arithmetic on those figures puts Cascade at seventy-six point eight on tokens, seventy-one point eight on time, and sixty-eight point zero on cost.

The cost saving outpaces the token saving — thirty-two percent against twenty-three point two — which means the efficiency gain is not purely a matter of using fewer tokens. The blog provides no breakdown of why that gap exists, no workload, no model specification, and no run count, so no explanation is offered here.

THE ACCOUNT LAYER BECOMES STRATEGIC SURFACE

Whoever controls the identity primitives for agents — the email namespace, the phone number pool, the wallet rails — sits between the agent and every external system it touches. That is a different competitive position from model quality or orchestration speed. It is closer to how enterprise identity providers (and corporate card networks) have historically accumulated leverage: quietly, through infrastructure, before the market recognised it as a moat.

DELEGATED SPENDING IS THE ONLY KIND THAT SHIPS TODAY

The wallet pays within the budget the user sets — capped and delegated, not autonomous. That constraint is not a limitation the company apologises for; it is the design. Where the bound sits, who sets it, and who carries the loss when an agent spends inside it are questions the blog does not address, and this piece does not answer them either.

EARLY ACCESS IS A SIGNAL, NOT A LAUNCH

Cue is invite-only, first-come first-served, with iOS still pending App Store review. That is a staged rollout designed to control load and gather signal before general availability — standard practice, but worth naming clearly given how launch announcements can blur into launch reality. The efficiency figures from Cascade apply to one tested configuration. How Cue performs across a broad, diverse user base under real-world conditions is a separate question the announcement does not answer.

The Bottom Line

Manus 2.0’s headline numbers — from one tested configuration, measured by Manus against its own previous system, with no independent verification — are less structurally interesting than the four words buried in the identity announcement: email, phone, wallet, computer. Those are not AI features. They are the account primitives that every external system uses to decide whether a counterparty is real and trustworthy. Issuing them to agents moves the competitive question from “how good is the model” to “who owns the trust surface between agents and the world” — and that is a harder question to catch up on once it is settled.

Source: Manus company blog, September 28, 2026. All efficiency figures are from that promotional document. No workload, model specification, run count, independent benchmark, user number, revenue figure, pricing, payment rail, or statement on real money versus credits is available from the source or supplied here. Nothing in this article is investment advice.