Manus AI Agent TESTED | First Impression

Manus AI Agent proves it can do end-to-end “agentic” workflows in a browser and a sandbox: it logs into X.com, researches a concept, drafts and ultimately posts a tweet, then builds a research-backed presentation with images, and finally turns an uploaded CSV of API pricing links into a report with generated graphs. The most consequential takeaway is not just that it can answer questions—it can operate tools (a live browser session, a file system, and code execution) to produce publishable outputs.

In the first test, Manus is given an explicit goal: log into x.com, research “vibe coding,” and post a tweet. It connects to a containerized/virtualized browser environment and asks for takeover to proceed. After logging into an experimental account without issue, it performs web research across multiple sources, compiles a short factual summary, and drafts tweet text. Manus then presents options for the user to choose from rather than posting immediately. When instructed again to post a selected option, it navigates the X interface, pastes the drafted text into the “What’s happening” box, and successfully posts. The result is functional but not flawless—posting required a workaround after earlier behavior suggested it might avoid direct posting due to built-in safety limits.

The second test shifts from social posting to document generation. Manus is tasked with creating a presentation on “Anthropic MCP servers,” including what they are, how to use them, and why they matter, plus images. It spins up a working directory, pulls documentation and reference material from Anthropic’s pages, and writes a set of research notes. From there, it generates a slide deck in MDX format (“Presentation.mdx”), assembling content sections such as key components, architecture, tool resources, prompts, and a conclusion/Q&A structure. Image handling is the main friction point: it briefly gets stuck searching for images (including an architecture diagram), but the run continues and produces a usable deliverable. The final output is described as “not perfect,” yet coherent and easy to follow, with a clear to-do-like progression during creation.

The final test evaluates file-driven analysis. Manus receives an uploaded CSV containing URLs to API pricing pages across providers (including Google, OpenAI, DeepSeek, and Anthropic). It browses the linked pages, extracts token pricing for multiple models, and writes a combined dataset. Some links are missed—DeepSeek pricing doesn’t get captured during the first pass—and OpenAI pricing encounters a Cloudflare verification barrier, limiting access to the newest figures. Even with those gaps, Manus proceeds by generating a Python script, installing missing plotting libraries in the sandbox (e.g., seaborn), and producing graphs and a report with embedded visuals. The deliverable includes input and output token comparisons across selected models (e.g., Claude 3.5 Sonnet, Claude 3.5 Haiku, Gemini 2.0 Flash, GPT-4 Turbo, and GPT-3.5 Turbo), with the report delivered via a generated URL.

Across three sessions, Manus looks capable of chaining browsing, research, structured writing, and code-based visualization into practical outputs. The performance is still uneven—access restrictions and image/pricing extraction issues show where agent tooling remains brittle—but the overall workflow readiness is strong enough to feel genuinely useful for iterative, real-world tasks.