Manus and the general-agent push: what a do-anything AI really automates

The demos are irresistible: an AI agent screens job candidates, builds a competitive analysis deck, books travel, and deploys a website—all from a single prompt, no human in the loop. Manus made that pitch in March 2025 and the internet lost its mind. A year on, with the product acquired by Meta and the general-agent category expanding fast, it’s worth asking a plainly operator-focused question: what does a “do-anything” agent actually get done reliably, and where does the loop break?

What Manus actually is (and what just happened to it)

Launched around March 2025, Manus’s core proposition marks a significant departure from conventional conversational AI: it is designed not merely to respond or suggest, but to independently plan, execute, and deliver results for complex, multi-step tasks across various domains.

The platform’s core idea is simple but ambitious: give an AI agent a full virtual computer—browser, terminal, file system—and let it complete multi-step tasks autonomously.

The acquisition story added a major plot twist. Meta Platforms acquired Manus, a Singapore-based developer of general-purpose AI agents, in late December 2025. Manus, founded in China before relocating to Singapore, had executed complex tasks such as market research, coding, and data analysis—and claimed an annualized average revenue of more than $100 million just eight months after launch.

The Wall Street Journal reported the deal closed at an amount over $2 billion. However, operators should note an unresolved complication: on April 27, 2026, China’s National Development and Reform Commission blocked Meta’s acquisition of Manus, requiring “the parties involved to withdraw the acquisition transaction.” The product continues to run subscriptions, but the corporate situation warrants monitoring before you build deep production dependencies on it.

What the general-agent pattern can actually automate

At the core, Manus—and tools like OpenAI’s browser agent or Anthropic Computer Use—share the same architecture: they take a high-level goal, decompose it into sub-tasks, and execute against a live browser and file system. Manus can manage complete workflows from start to finish: research topics, navigate websites, fill forms, generate content, and even run or deploy code, all without constant user guidance.

Where operators consistently find real ROI:

Open-ended web research. Give it a research question with clear deliverable expectations (a structured report, a spreadsheet, a comparison table) and general-purpose agents perform well. Based on what users are sharing, it can code simple games, generate working web pages, create images, and do some deep research.
Browser tasks with no API. For the last decade, SME automation progressed only where APIs existed. The 60% of work that requires “logging into a website and clicking through things” has been an automation black hole. AI browser agents close that last mile.
Structured file output. Tasks with clear terminal artifacts—a .docx, a slide deck, a CSV—outperform tasks that require the agent to “know when it’s done.” Everything happens in the cloud, so you don’t need to keep your computer on. Once you start a task, Manus keeps working until it is done, and then it sends you a message with the results.

Where the loop breaks: honest limitations

The capability is real. The reliability gap is also real, and that gap matters operationally.

Error compounding at scale. The idea of fully autonomous multi-step agents sounds great, but in practice it falls apart under simple math. The issue isn’t intelligence or prompt quality—it’s compounded error rates. Even small per-step mistakes grow exponentially over time, which makes true end-to-end autonomy impossible at scale. The International AI Safety Report found the same: some agents can plan and complete multi-step tasks over extended time horizons, albeit with limitations on reliability and largely in controlled environments.

The “permission loop” problem. OpenAI’s browser agent experience is instructive here. The core problem with Operator, showing up in every honest review, is that it stops constantly to ask for permission. It hits a login screen and pauses. It encounters a dropdown it’s not sure about and asks you to confirm. It reaches a checkout page and wants to double-check before it does anything.

Operator’s headline use case—book a flight, order a product—had a sub-50 percent success rate on real-world sites , which contributed to OpenAI shutting it down as a standalone product in August 2025 and absorbing it into ChatGPT Agent.

No persistent state. The limitation is that nothing persists. Every script runs in a disposable VM. There is no package manager state, no installed libraries carried between sessions, and no version control. For one-shot research jobs this is fine. For recurring production workflows, it’s a blocker.

Output quality is inconsistent. The quality is a mixed bag. The output often lacks polish. A generated webpage might have a wild color scheme, or the code might have small bugs that a developer has to fix anyway.

The pricing reality—and the unpredictability tax

Instead of a simple monthly subscription, Manus runs on a “credit” system. Every single thing the agent does—from writing one line of code to creating one slide—eats up a certain number of credits. The biggest problem people are running into is how unpredictable this model is.

Users have reported that a single, moderately complex task can blow through over 900 credits. Some even used up their 1,000 free starter credits on their very first request.

Manus’s Pro plan starts at $20/month and goes up to $200/month. The free tier gives you 300 daily refresh credits to test the waters. But operators should budget conservatively: Manus doesn’t tell you how many credits a task will cost before you hit go. You’re essentially rolling the dice every time, with no clue if your request will cost a few pennies or a huge chunk of your monthly budget.

For teams with heavier workloads, Manus also offers dedicated Team and Enterprise plans with flexible per-seat pricing, centralized billing, and enterprise-grade security features, available by contacting Manus directly and designed for organizations that need administrative controls, SSO, and compliance features.

How to scope general agents for real workflows

The operators who report the best results with general agents aren’t using them as a replacement for structured automation—they’re slotting them into gaps where structured automation was never feasible. A few principles that hold up across the teams we talk to:

Constrain the task surface. Prompt specificity is leverage. Clear, well-defined tasks help Manus AI work more efficiently, reducing unnecessary iterations and credit consumption. “Research AI” is less efficient than “Research the top 5 AI trends in healthcare for 2024 and create a summary report.” Simple questions don’t require the full autonomous agent—use Chat mode for quick queries to save credits.

Human-in-the-loop on irreversible actions. Require human confirmation for any action above an amount threshold, any email send, or any critical data modification. The agent’s autonomy is a feature on research tasks and a liability on write operations.

Stack against purpose-built tools for structured work. General agents complement—not replace—dedicated workflow tools. If you’re running outbound sequences or automating multi-step ops, tools like Lindy or n8n give you durable, auditable flows that a general agent isn’t designed to maintain. Similarly, Browser Use as an open-source layer gives developers more control over the exact browser-automation surface than a managed product like Manus does.

Watch the benchmark math. Real performance still trails humans—Operator at 38.1% OSWorld vs. ~72% human baseline.

The real winners will build focused, domain-specific assistants that use AI where it helps most, but still rely on humans or deterministic systems for critical control points.

The broader category context

Manus isn’t operating in a vacuum. OpenAI integrated Operator into ChatGPT as “ChatGPT agent,” accessible by selecting “agent mode” from the dropdown in the composer. Anthropic’s computer use capabilities sit in the API for builders who want to go lower-level. Compare the tradeoffs across OpenAI Operator vs. Perplexity Comet to get a sense of how the browser-agent field is fragmenting by architecture. Computer use is becoming table stakes in frontier AI products, not a standalone offering.

Bottom line: General autonomous agents like Manus are genuinely useful for open-ended research, browser tasks that have no API, and one-shot deliverable creation—but they remain unreliable for chained production workflows where compounding errors, permission interruptions, and session-state loss create real operational drag. The credit-based pricing model adds a layer of cost unpredictability that makes budgeting difficult at scale. Treat them as a powerful but supervised layer in your stack, not a replacement for purpose-built automation—and keep an eye on the Manus/Meta acquisition situation before committing them to critical infrastructure.