Technical Context
I look at the Manus situation pragmatically: the fact of Meta's acquisition in late December 2025 is confirmed, but the degradation of quality, token "devouring," and broken billing are currently user signals, not backed by public metrics or official status pages.
But it's exactly these signals I always consider in AI architecture: SaaS AI has two pain points — billing and predictable consumption. If a user writes that "a monthly token supply vanished on who knows what," I read this as a lack of transparent telemetry: there is no clear breakdown by tasks, agent steps, retries, tool errors, and repeated model calls.
The second red zone is the payment flow. Broken billing means more than just inconvenience; it is a risk of halting workflows and the inability to quickly purchase limits at a critical moment.
I should separately note: Manus was historically perceived as an autonomous agent for complex tasks (research, code, data analysis, artifact collection). Therefore, its attempts to be used as a "swiss army knife" for presentations are logical, but it's precisely in such scenarios that the cost of errors is maximal: the agent can repeatedly regenerate slides, pull external tools, and burn through the budget without a clear result.
Business Impact and AI Automation
When a product enters the orbit of a major platform, I always account for a period of turbulence: account migrations, new data policies, infrastructure rebuilding, and roadmap priority shifts. If bugs occur alongside this, the business pays twice — in money and downtime.
Who wins? Teams whose AI implementation is not reliant "on a single service," but operates through an orchestration layer: task routing, limits, result caching, retry control, and logging. Who loses? Those who tied presentations, reports, and analytics to a single agent without a backup plan and without token consumption control.
In Nahornyi AI Lab projects, I usually set a simple rule: any external AI tool must be replaceable within 1–3 days. This is achieved not by magic, but by discipline: a unified interface for calling models/agents, prompt isolation, version control, and a separate billing module with limits per user, task, and day.
What should be done with the presentation task if Manus is "storming"? I don't take the comment "should have used Kimi Slides" as proof of Kimi Slides' superiority — I don't have confirmed data on the product. But the strategy itself is correct: keep a specialized tool for slides separate from an autonomous agent that might drift into expensive iterations.
Strategic Vision and Deep Analysis
My forecast for 2026: the market will split into two classes of solutions. The first — platform agents "for everything," which are great with integrations but unpredictable in terms of changes and policies. The second — narrow tools (presentations, sales, support), which are easier to stabilize and measure.
If you are building AI solutions for business, I wouldn't bet on a "universal agent" as the single source of truth. I would design the AI solution architecture so that presentations live in a pipeline: data → structure → slides → QA → export, and at each step, you can change the provider (Manus, another agent, a separate slide generator) without rewriting the whole process.
From Nahornyi AI Lab's practice: the most expensive incidents occur not because of text quality, but due to a lack of guardrails. Agent step limits, external call restrictions, bans on infinite "improvements," and mandatory consumption reports are not options, but basic AI integration for the real sector.
If complaints about Manus are confirmed and become widespread, companies with such an architecture will simply switch to alternatives. Companies without it will "fix the process" through manual labor and lose momentum.
This analysis was prepared by Vadym Nahornyi — lead expert at Nahornyi AI Lab on artificial intelligence implementation and AI automation in the real sector. I invite you to discuss your case: I will assess the risks of your current AI tools, design a replaceable scheme (orchestration, billing, observability), and help create AI automation that doesn't break due to a single external SaaS.