Everything About AI
NextFin News - Manus is betting that a general-purpose artificial-intelligence agent — software that takes a vague goal and returns a finished result without step-by-step supervision — can still win a market that is rapidly being fenced off by bigger, better-funded rivals. On September 1, the Singapore-headquartered company said it had formally resumed independent operations after unwinding its roughly $2 billion acquisition by Meta Platforms, and it signaled a fresh wave of product releases aimed at embedding its agent deeper into everyday workflows. The announcement lands as the agent category tips from novelty into infrastructure: analysts project the global AI-agent market to reach between $10.9 billion and $12 billion in 2026 and roughly $50 billion by 2030, a compound annual growth rate above 44 percent.
The question Manus's renewed push forces investors and competitors to answer is not whether agents are the next platform shift — that battle is largely settled. It is whether a standalone general agent can survive between two pincers: the model-layer giants bundling agents into products billions of people already use, and the vertical specialists that do one job better than a generalist ever could.
The Situation: Independence, Then Expansion
Manus first captured attention in March 2025 when it launched its first general AI agent, a system designed to plan, browse the web, write and run code, and assemble finished artifacts — a research report, a website, a slide deck — with minimal human hand-holding. In a statement about Meta's acquisition, the company said its system had processed more than 147 trillion tokens and powered more than 80 million virtual computers, a scale claim that framed the agent as infrastructure rather than a toy. Growth followed quickly: Manus said it crossed $100 million in annual recurring revenue eight months after launch, with a total revenue run-rate above $125 million and month-over-month growth exceeding 20 percent since the release of Manus 1.5.
That trajectory made the company an acquisition target. Meta Platforms agreed to buy Manus, with the deal closing on December 29, 2025. The logic was distribution: Meta's family of apps reaches billions of users, and Manus's agent could theoretically be woven into Meta AI, Facebook, Instagram and WhatsApp. But the deal did not survive regulatory scrutiny. Beijing blocked the transaction, and on August 11 Manus said it would resume "operating as an independent company." The unwind carried real operational cost: data generated by affected users on or after the December 29 acquisition date was scheduled for deletion between August 23 and August 24, with a restoration portal opening August 25.
"Manus has formally resumed independent operations. Our founding team will continue to lead the company, with a relentless commitment to product innovation and advancing general AI agents for our users worldwide," the company said in a September 1 post. "Next, you will see Manus become more deeply embedded in your everyday workflows, interact more directly with the world around you, and act more proactively on your behalf."
The phrase "more proactively on your behalf" is the operative one. It points away from the reactive chatbot model — type a prompt, get an answer — and toward agents that initiate work: monitoring calendars, watching data, triggering tasks without being asked. That is the frontier the category is moving toward, and Manus is trying to reach it before the market consolidates around incumbents.
Why the Agent Market Is Not a Normal Software Market
To understand Manus's position, it helps to separate what is cyclical from what is structural in the current agent boom. The cyclical layer is the hype cycle itself: funding rounds, valuation multiples, and the rush of enterprises running pilots that never reach production. One widely cited estimate holds that 40 percent of enterprise applications will embed task-specific AI agents by the end of 2026, up from less than 5 percent in 2025 — but other surveys put actual production deployment far lower, around 11 percent of enterprises that have adopted agents in some form. That gap between adoption and production is the classic shape of a technology spending wave that will partially mean-revert: pilots get cut, budgets tighten, and the weakest vendors disappear.
The structural layer is different, and it is why the biggest technology companies are treating agents as an existential priority rather than a feature line. For forty years, software has followed the same contract: the human supplies intent step by step, and the machine executes each instruction. Agents invert that contract. The human supplies an outcome — "plan a trip," "build a landing page," "analyze this quarter's churn" — and the machine decomposes the goal, calls tools, recovers from errors, and delivers the artifact. That inversion changes where value accrues. It moves the center of gravity from the interface and from individual applications toward the action layer: the authenticated connections to email, calendars, code repositories, browsers, payment systems and enterprise software that an agent must control to complete real work.
This is why the market is growing at a 44 to 46 percent annual clip toward a projected $50 billion by 2030, and why the number is almost certainly understated. Agent value does not sit in a single product category. It bleeds into search, productivity, customer service, software development and business-process automation. A chatbot charges per seat. An agent that completes a task can eventually charge per outcome — and outcome-based pricing, already gaining traction in 2026, expands the revenue pool faster than seat counts would suggest.
The Two Pincers Closing on a General Agent
Manus's strategic problem is that it occupies the most contested real estate in the agent economy. It is a general-purpose agent, which means it competes on two fronts at once.
On the first front are the model-layer giants. OpenAI ships Operator, an agent that can act in a web browser on a user's behalf, and has folded agent capabilities into the broader ChatGPT experience. Anthropic, whose February 2026 funding round valued the company at $380 billion, built Claude Code and the Model Context Protocol, an open standard for connecting models to tools and data. Google, Microsoft and Amazon are pursuing the same logic: agents bundled into products that already have distribution, billing relationships and enterprise trust. Their advantage is not necessarily better agent technology. It is that they do not have to persuade a user to adopt a new product; the agent arrives inside a workflow the user already occupies.
On the second front are the specialists. Cursor, the AI-first coding environment, reached a $3 billion valuation in May 2026 and has projected revenue above $6 billion by year-end. For a developer, a coding-native agent is simply more fluent at the specific job of writing and editing code than a general agent is likely to be. The same logic applies in other verticals: a legal-research agent, a sales-automation agent, a data-pipeline agent. Each can optimize for one domain's workflows, data and failure modes in ways a horizontal generalist cannot.
The pincer is not theoretical. It is the same dynamic that compressed the standalone app layer in earlier platform shifts. When a capability becomes valuable enough, the platform bundles it; when a job is complex enough, a specialist out-executes the generalist. The middle thins out.
The Second-Order Question Nobody Is Asking
The consensus read of the agent race is straightforward: the best model wins. Manus's renewed push implies a different hypothesis, and it is worth taking seriously. The second-order question is not which company has the smartest frontier model. It is which company controls the action layer — the identity, permissions and authenticated connections that let an agent actually do things in the world.
An agent without tool access is a chatbot with a plan. To book a hotel, file an expense report, merge a pull request or reconcile a ledger, the agent must be logged in somewhere, with credentials, under a permission model the user trusts. That is a distribution and trust problem, not a model-quality problem. And it is the one place where a standalone agent can still build a defensible position: by accumulating the integrations, the workflow memory and the reliability record that make users willing to hand it credentials in the first place.
Meta's aborted acquisition is instructive here. The strategic prize was never Manus's model; it was the combination of Manus's agent with Meta's distribution and commerce graph. The deal's collapse returns Manus to independence, but it also returns Meta to square one on the action layer. Manus keeps its product and its paying users but loses the distribution engine that could have made it ubiquitous. Both sides now have to rebuild the missing half of the thesis.
This reframes what to watch. Model benchmarks will move every quarter and prove little about who wins. The durable signal is integration depth and task completion: how many real systems the agent can act in, how often it completes a multi-step task without human rescue, and whether users escalate from occasional use to daily reliance. That is where Manus's promised shift toward proactive, embedded workflows will be won or lost.
The Counter-Thesis: General Agents Are the Losing Middle
The strongest case against Manus is that the general-purpose agent is a transitional form, not a destination. The counter-argument runs like this: users will not maintain a separate agent subscription when the assistant they already pay for — ChatGPT Plus, Microsoft Copilot, Google's Gemini — can perform the same tasks. Bundling has won every comparable software category, from browsers to office suites to video conferencing. Meanwhile, for high-stakes work, specialists will dominate because they can guarantee reliability in a narrow domain. A general agent that is "good enough" at everything is priced toward zero in the middle.
This view has real evidence behind it. Cursor's growth shows developers choosing a domain-native tool over a general coding agent. Enterprise surveys show production deployments concentrated in narrow, high-return workflows rather than open-ended general agents. And the economics of agent inference are punishing: every failed task still consumes compute, and credit-based pricing — Manus's own model, with plans reported in the $19 to $199 per month range — creates friction when users cannot predict what a task will cost.
The counter-thesis is not easily dismissed, and it defines the falsifying signal for Manus's strategy. If, by the end of 2026, bundled agents from the model-layer giants are capturing the majority of new agent usage while Manus's paid-seat growth stalls, the standalone general-agent thesis fails. Conversely, if Manus can demonstrate that users escalate to daily, proactive use — that the agent becomes the first thing they open rather than a fallback — then the general agent survives as a category, and Manus's independence becomes an advantage rather than a handicap. The company's next product releases, promised as pushing "the boundaries of what general AI agents can do," will be the first real test.
What Comes Next
Three signals will separate narrative from traction over the next two quarters. First, product substance: Manus has promised embedded, proactive workflows, and the market will judge whether those are genuine capabilities or marketing language. Second, retention and escalation: revenue growth matters, but the ratio of daily active users to paid seats — and the share of tasks completed without human intervention — will matter more. Third, competitive response: every move Manus makes will be matched by OpenAI, Anthropic, Google and Microsoft inside products with far larger installed bases.
The base case is that the agent market keeps growing at a 40-plus percent clip and Manus remains a meaningful independent player with a loyal user base, but cedes mass-market share to bundled agents. The upside case is that proactive, embedded workflows prove to be a different product category altogether — one where a focused agent with deep integrations beats a bundled generalist — and Manus's early lead in general-agent usage compounds into a defensible action-layer position. The downside case is the pincer closes as expected: bundling captures casual users, specialists capture high-value workflows, and standalone general agents compress into a niche.
The agent economy is being built on a simple premise: software should take intent and return outcomes. Manus's renewed push is a bet that users will trust an independent company with that contract. The market is large enough, and shifting fast enough, that the bet is not irrational. But the window to prove it is narrowing — and the companies on the other side of the pincer are not waiting.
Explore more exclusive insights at nextfin.ai.
