AI Agent (5) — Manus: The Autonomous Agent That Plans and Executes Multi-Step Jobs

Learn what Manus is, how to choose between Chat and Agent Mode, how to write prompts that produce real artifacts, and how to combine Manus with other AI agents.

AI Agent (5) — Manus: The Autonomous Agent That Plans and Executes Multi-Step Jobs

In this guide, you will learn what Manus is and how it differs from Claude Skills and Perplexity Computer, how to choose between Chat Mode and Agent Mode, what kinds of long-horizon multi-step tasks Manus is built for (and which tasks it isn't), and how to write prompts that produce real, finished outputs — websites, slide decks, research reports — rather than generic summaries.

Difficulty: ★★★☆☆ (Approachable for non-developers, but the autonomy requires careful prompt-writing)
Required Tools: Manus account (free tier with 300 daily credits, or Pro from $20/mo with 4,000 monthly credits)
Updated: May 2026

Overview

If Claude Skills are specialists you call by name and Perplexity Computer is a research analyst that browses for you, Manus is the autonomous contractor you hand a goal to and walk away from. You tell Manus what you want — "build me a landing page for my pet-sitting startup with thoughtful copy and a clean design", or "research the top 8 enterprise CRMs and produce a 30-slide investor pitch comparing them" — and Manus plans the steps, browses the web, writes the code, generates the design, executes the work, and delivers a real downloadable artifact: a working website, a formatted PowerPoint deck, an Excel spreadsheet with charts, a Word document. The outputs are fully editable and yours to keep.

What makes Manus distinctive in 2026: it's a multi-agent system. Most AI tools use one large language model. Manus orchestrates several specialized sub-agents — a browser agent for web navigation, a code agent for writing and running scripts, a file-system agent for managing outputs, a model-picker that routes each step to the best LLM for the job (Claude, GPT-5.5, Qwen, others). This architecture is what lets Manus handle long-horizon tasks that confuse single-LLM agents — tasks with 20, 30, or 50 steps where context maintenance breaks single-model approaches.

Manus was originally built by the Monica.im team in China and was acquired by Meta in late 2025 for approximately $2 billion. Since the acquisition, the platform has expanded beyond research and writing into a full production environment with a Web App Builder, AI-powered slide creation, a desktop app with local file access, and native integrations with Slack, WhatsApp, and Telegram. In this article, we'll cover the two modes (Chat and Agent), the kinds of tasks Manus handles brilliantly, the tasks where it's overkill, and how to write prompts that get genuine outputs rather than apologetic stubs.

Who This Is Useful For

People with long-horizon tasks — research projects, content production, app or site builds — that take hours of manual work and don't fit neatly into a fixed Zapier flow

Founders and solopreneurs who need real outputs (a working landing page, a pitch deck, a market research report) without paying a designer, developer, or analyst for each one

Anyone curious about agents but priced out of Perplexity Max ($200/mo) — Manus's free tier and $20/mo Pro plan offer real autonomous agent capabilities at a much lower entry point

Teams managing many parallel projects who could benefit from running several Manus tasks simultaneously in the background while focusing on something else

What You Will Learn

By the end of this article, you'll be able to do five things:

Distinguish between Manus's Chat Mode and Agent Mode — and pick the right one for any task

Identify which tasks Manus is genuinely built for versus which tasks belong to other agents (Claude Skills, Perplexity Computer) or just regular chat

Write prompts specific enough to produce real, finished outputs — not generic stubs that need rebuilding

Manage credits without surprise bill shock — understanding which tasks burn 30 credits versus 500 credits

Combine Manus with your other AI tools — Claude for fine-tuning, Perplexity for live research, Manus for execution-heavy production

What You Need

A Manus account — free tier requires no credit card and gives you 300 daily refresh credits with up to 5 concurrent tasks

About 25 minutes to read and 30+ minutes to run your first real task (Manus tasks often run autonomously for 15–60 minutes)

One real long-horizon project in mind — a website you've been meaning to build, a research report you've been avoiding, a slide deck you'd otherwise hire someone for

Step 1 — What Manus Actually Is (And Isn't)

The clearest way to understand Manus is by what it produces. Most AI tools produce text — a chat reply, a research summary, a list of suggestions. Manus produces artifacts: a deployed website with a public URL, a downloadable PowerPoint file with charts, a fully formatted Word document, an Excel spreadsheet with formulas and pivot tables, even a working web app. The output is an actual file or live thing you can use, share, or edit.

Three things make this possible:

A virtual computer environment. Manus runs in a sandboxed environment with a real browser, a real terminal, a real file system, and the ability to execute code (Python, JavaScript, shell commands). When Manus says "I'll build a landing page," it actually opens a code editor in its environment, writes HTML/CSS/JavaScript, runs a build, and deploys it.

Multi-agent orchestration. A single LLM tends to drift on tasks longer than 8–10 steps. Manus breaks each task into sub-tasks and assigns them to specialist sub-agents — browsing, code-writing, design, data analysis. Each sub-agent reports back to a coordinator that maintains the overall plan. This architecture is why Manus can handle 30+ step jobs without losing the thread.

Smart model routing. Each sub-task gets the right model for the job. Quick browsing might use a fast cheap model; technical synthesis might use Claude Opus or GPT-5.5; data-heavy work might use Qwen. You don't pick the model — Manus picks it for you, optimizing for both quality and cost.

What Manus is not:

Not a chat interface for quick questions. Reaching for Manus for a 2-sentence question wastes credits and time. Use ChatGPT, Claude, or Perplexity for chat-level work.

Not a fixed automation tool. Manus decides what to do based on the goal you give it. If you want a predictable, repeatable pipeline that runs the same way every time, use Zapier or Make instead.

Not a substitute for human judgment on high-stakes decisions. Manus can produce a research report on which startup to invest in, but it can't make the decision for you. The output is a starting point, not a verdict.

Not always cheaper than human work. A complex Manus task can burn 200–500 credits ($5–15 of consumed budget on Pro plan). For some tasks, that's cheaper than a freelancer; for trivial tasks, it's more expensive than just doing it yourself.

Step 2 — Chat Mode vs Agent Mode

Manus has two operating modes, and picking the right one is the single biggest determinant of whether you get a useful result.

Chat Mode. Chat Mode behaves like a smarter version of ChatGPT — you ask, Manus answers. The conversation is the work. Chat Mode uses fewer credits (because there's less orchestration overhead), runs faster, and is the right choice for one-shot questions, brainstorming, and quick text outputs. Use Chat Mode when:

You're asking a single question with a single answer

You want to draft, edit, or summarize text

You're testing whether Manus understands a topic before launching a full Agent task

You want fast (under 30 seconds) replies, not long autonomous runs

Agent Mode. Agent Mode is the autonomous, multi-step, real-output mode that Manus is famous for. You give it a goal; Manus plans the steps, executes them across browser/terminal/files, and produces a real artifact. Agent Mode tasks typically run for 15–60 minutes (some longer); you can leave them running and come back. Use Agent Mode when:

The task has multiple distinct phases (research → analyze → produce → format → export)

The output is a real artifact (website, slide deck, document, data analysis)

You want Manus to actually browse, run code, or generate files — not just respond with text

The task takes you 1+ hours of focused work to do manually

The default mistake new users make is asking complex multi-step tasks in Chat Mode and getting a brief text reply instead of a working artifact. Switch to Agent Mode for any task where you want Manus to actually do something rather than describe what could be done.

Step 3 — Plans, Pricing, and Real Credit Costs

Manus's pricing is simpler than most agent platforms, but the credit math is worth understanding before you start.

Plan	Price	Daily/Monthly Credits	Concurrent Tasks
Free	$0	300 daily refresh credits	5
Pro	$20/mo or $192/year	4,000 monthly credits	20
Pro+	Higher tier	Larger credit pool, more concurrent tasks	Higher
Annual billing		17% discount (~2 months free per year)	Same

How credits actually burn. Credits aren't tied to time — they're tied to complexity and tool use. A rough mental model:

A simple Chat Mode reply uses 1–5 credits — basically free

A short Agent task (5-10 minute run, single output like a 1-page brief) typically uses 20–60 credits

A medium Agent task (15-30 min run, web app or 10-slide deck) typically uses 80–200 credits

A long Agent task (45+ min, deep research with 30-slide deck or full website build) typically uses 200–500 credits

An extreme task (multi-hour, hundreds of sources, complex outputs) can use 500–1,500 credits

Free tier (300 daily credits) covers roughly 1–3 medium Agent tasks per day. Pro at $20/mo (4,000 monthly credits) covers roughly 20–40 medium Agent tasks per month, comfortable for most active users.

Step 4 — Five Real Tasks Manus Excels At

Task 1: Build a landing page or simple website. Tell Manus the product, audience, and design vibe; Manus writes the HTML/CSS/JavaScript, generates copy, picks images (or generates them), and deploys to a public URL. "Create a landing page for a time-tracking SaaS app. Make the design unique — not like every other AI-generated site. Include a hero section, three feature cards, a pricing table with three tiers, and an email capture form." Result: a real working page in 20–30 minutes, downloadable code or live URL.

Task 2: Generate a professional slide deck. Manus's slide generation is one of its strongest features — it researches the topic first, builds an outline, then generates each slide with appropriate design. "Create a 20-slide investor pitch deck for my pet-sitting startup. Cover problem, market size, solution, traction, team, financial projections, ask. Match a clean modern design." Result: an editable .pptx file with charts, images, and proper structure.

Task 3: Run a multi-source research project. Like Perplexity Computer but with deeper synthesis and the ability to produce richer outputs (charts, tables, full reports). "Research the Asian electric scooter market in 2026. Compare the top 10 brands across price, range, weight, smart features, and warranty. Produce a 30-page report with charts." Result: a downloadable PDF or Word document with embedded data analysis.

Task 4: Analyze a dataset and produce visualizations. Upload a CSV or point Manus at a dataset; Manus runs Python code to analyze it, finds patterns, generates charts, and writes a summary. "Pull the last 12 months of sales data from this CSV. Find seasonal patterns. Suggest pricing changes. Produce charts and a 5-page write-up." Result: an Excel file with analysis sheets plus a Word summary, both ready to share.

Task 5: Connected workflow tasks across messaging. With Slack, WhatsApp, or Telegram integrations enabled, Manus can monitor channels, triage messages, draft replies, or escalate based on rules. "In my customer support Slack channel, watch for messages mentioning refunds. For each one, look up the customer's order history in our system, draft a reply suggesting next steps, and post the draft for me to review." Result: ongoing autonomous triage that you supervise rather than execute.

Step 5 — Writing Prompts That Get Real Output

The biggest difference between a Manus task that burns 200 credits and produces a usable artifact, versus one that burns 200 credits and produces a vague apology, is the specificity of the initial prompt. Manus respects detail more than most AI tools. The four-section formula:


[The specific deliverable in 1 sentence — name what you want to walk away with]Goal:
[Why you're building this; who it's for]
Specifications:
Format / type: [website / slide deck / report / spreadsheet / etc.]

Length / size: [number of pages, slides, sections]

Audience: [who will read or use this]

Style / tone: [be specific — give 2-3 reference adjectives or sites you like]
Content:
[List the actual content sections, points, or data Manus should include — be detailed]
Constraints:
[What NOT to include]

[Any time, budget, or scope limits]
Output format:
[Exactly which file types you want — .pptx, .docx, .pdf, .xlsx, deployed website URL, etc.]

[Where to deliver — download link, email, save to a folder]
If you encounter a question or ambiguity, [ask me before proceeding / make your best judgment and flag it / stop and wait for input].

Each section earns its place:

The deliverable line forces specificity at the start. "A pitch deck" is vague; "a 20-slide investor pitch deck" is concrete. Manus's plans get measurably better when you name the artifact.

The Goal section keeps Manus aligned with your reality. Without it, Manus produces generic outputs. With it, every step gets filtered through "does this serve the actual goal?"

Specifications lock down format, audience, and tone. The single biggest reason for unusable Manus output is missing specifications — Manus had to guess on tone or length and guessed wrong.

Content is where you give Manus your subject-matter expertise. The more you specify, the less Manus has to invent (and the less it hallucinates).

Constraints are underused. "Don't include stock photos." "Limit to under 30 minutes runtime." "Don't generate more than 25 slides." These prevent over-engineering.

Output format is critical. Manus can produce .pptx, .docx, .xlsx, .pdf, deployed websites, ZIP files, and more. Naming the format upfront saves you from getting an HTML preview when you wanted a slide deck.

The "if you encounter ambiguity" instruction is one of the most useful patterns in long-task prompts. Most users want Manus to ask before going off-script; saying so explicitly improves outcomes dramatically.

Step 6 — A Real Walkthrough: Build a Pitch Deck for Your Side Project

Let's run a real Agent Mode task. We'll build a pitch deck for a hypothetical side project — adjust the prompt to match yours.

Open Manus, switch to Agent Mode, and paste this prompt:

Build me a 20-slide investor pitch deck for my pet-sitting startup, "PawSit".Goal: I'm pitching to angel investors and small VCs in Singapore and Taipei Looking to raise SGD 250K seed round The audience knows the SEA pet market loosely; assume they understand startups but not pet care specifics Specifications: Format: PowerPoint .pptx file, 20 slides, 16:9 aspect ratio Style: clean modern minimalist design (think Linear's branding), use teal and warm gray as primary colors Tone: confident but not aggressive, data-led, no buzzwords Length: ~30 seconds per slide when presented = 10-minute pitch Content (use this slide structure): 1. Cover (PawSit + tagline + my name) 2. Problem — pet owners in SEA struggle with reliable trusted sitters 3. Market size — SEA pet care market data, ideally with chart 4. Solution — PawSit overview 5. How it works — 3-step user flow 6. Product screenshots (use placeholder boxes I'll replace) 7-9. Key features — vetting, insurance, trust system 10. Traction — placeholder for our current numbers 11. Business model — commission take rates 12. Unit economics — back-of-envelope LTV/CAC 13. Competition — vs DogVacay, Wag, regional players 14. Why now — timing thesis 15. Team — placeholder for me + co-founder 16. Roadmap — next 18 months 17. Financial projections — Y1-Y3 (use realistic SaaS-like numbers) 18. The ask — SGD 250K, use of funds breakdown 19. Vision — where we are in 5 years 20. Thank you + contact Content I want filled in by you: Real-ish market size figures (cite source pages) Realistic competitive positioning Standard SaaS-style financial projection ranges Content I will fill in myself: Team bios Real traction numbers Real unit economics Specific use-of-funds amounts Constraints: No stock photos No charts using fake data — only use real cited data or clearly labeled placeholder data Don't make slides text-heavy (max 4 bullets per slide) Do NOT exceed 20 slides Output format: Final .pptx download link Also save a Google Slides version if possible

If you encounter a question or ambiguity, ask me before proceeding rather than guessing.

Click Run in Agent Mode. Manus shows you its plan first — a numbered list of steps it will take (research the SEA pet market, design the slide template, populate slides 1-6 with content, etc.). Review the plan; if it looks right, approve it.

Manus runs autonomously for 25–45 minutes. While it runs, you can:

Continue using Manus for other tasks (up to 5 concurrent on free tier, 20 on Pro)

Switch to other apps and let Manus work in the background

Periodically check progress in the Manus task panel

Receive a notification when the task completes

When the task finishes, you'll get:

A download link for the final .pptx file

(If requested) a Google Slides URL

A summary of what Manus did, including the sources it used for market data

Any questions Manus had to make judgment calls on, flagged so you can review

Open the .pptx in PowerPoint or Keynote. You'll typically need to make 5–10 manual edits — fixing the team bios, plugging in your real numbers, refining one or two slides where Manus's interpretation didn't quite match yours. But the heavy lifting (structure, design, market research, financial framing) is done.

Step 7 — When Manus Is the Wrong Tool

Three categories of task where reaching for Manus wastes time and credits.

Category 1: Quick chat-level questions. "What's the difference between SaaS and PaaS?" doesn't need an autonomous multi-agent system. Use ChatGPT or Claude in regular chat — it'll be done in 2 seconds and cost nothing.

Category 2: Predictable repeatable workflows. If you do the same thing the same way every week with predictable inputs, build a Zapier or Make automation instead. Manus's strength is adapting to fuzzy goals; that strength becomes overhead when the workflow is fixed.

Category 3: Tasks where wrong output is high-stakes. Manus is a generative agent — it can produce confident-sounding wrong answers. For tasks where the cost of a wrong output is high (legal contracts, medical advice, financial filings, regulatory compliance), use Manus to draft and a qualified human to verify. Don't trust autonomous output for high-stakes work without explicit human review.

Step 8 — Combining Manus With Your Other AI Tools

Manus is most powerful when paired with the other agents from this series.

Manus + Perplexity Computer. Perplexity Computer specializes in web research with strong citation discipline. Manus specializes in producing real artifacts. Use Perplexity to do the research phase ("compare 10 mortgage products with sources"), copy the report into a Manus prompt, and have Manus build the final artifact ("turn this comparison into a client-ready 15-page report"). Perplexity does the gathering; Manus does the production.

Manus + Claude Skills. Manus produces; Claude polishes. After Manus generates a slide deck or research document, run it through a Claude Skill like /polish-for-investor-tone to refine the voice and tighten the language. Manus is great at structure and content; Claude is great at finishing touches.

Manus + Cowork. Manus produces an artifact; Cowork files it locally and integrates with your other documents. Manus generates a Word report; Cowork saves it into your project folder, organizes it with related files, and updates a master index document. Two specialists, clean handoff.

Manus alone for end-to-end production. When the whole pipeline (research → analysis → production → export) needs to happen without your supervision, Manus's multi-agent architecture handles the whole thing in one task. This is when Manus genuinely earns its credits versus stitching together multiple tools.

Common Mistakes to Avoid

Three patterns that consistently waste credits or produce poor outputs.

Mistake #1: Vague prompts in Agent Mode. "Make me a website" or "Research the AI market" produces output that's either too generic to use or so broad that Manus burns hundreds of credits on irrelevant work. The 7-section prompt formula in Step 5 is worth memorizing — the discipline pays for itself in the first task.

Mistake #2: Skipping the plan-review step. Manus offers to show you its plan before executing for any non-trivial task. Most users skip this and run blind. The 30 seconds spent reviewing the plan is the cheapest insurance against 200-credit mistakes.

Mistake #3: Running Manus for tasks chat would handle. Quick questions, simple summaries, single-paragraph drafts — these belong in regular chat (ChatGPT, Claude, free Perplexity). Reaching for Manus's full Agent Mode for trivial tasks is like flying a private jet to the corner store. Save Manus for tasks that genuinely need its capabilities.

Going Further

Start with a free-tier exploration week. Sign up for the free Manus tier and run 3–5 real tasks. Pay attention to how the credit math works in practice and how often the autonomous output meets your bar. Most people learn enough in a week to know whether Pro at $20/mo would pay back.

Build a personal "Manus playbook". Keep a Notion or document with your best Manus prompts categorized by output type — landing page prompt, slide deck prompt, research report prompt, data analysis prompt. Once you have 5–10 working prompts, your future tasks become "edit existing prompt + paste new context" rather than "write a new prompt from scratch."

Use Manus's messaging integrations cautiously. Slack/WhatsApp/Telegram integrations are powerful but also high-risk for autonomous mistakes (wrong recipient, wrong timing). Start in monitoring or draft mode (Manus surfaces drafts for your approval) before letting Manus actually send anything on your behalf.

Key Takeaways

Here's what you learned in this guide:

Manus is the autonomous contractor of the agent landscape. You give it a goal; it produces real artifacts (websites, slide decks, reports, spreadsheets) using a multi-agent architecture and smart model routing.

Two modes — pick the right one. Chat Mode for quick questions and drafts. Agent Mode for multi-step autonomous production tasks.

Pricing is approachable. Free tier with 300 daily credits is enough for 1–3 medium tasks per day. Pro at $20/mo with 4,000 monthly credits covers most active users.

The 7-section prompt formula transforms output quality. Specific deliverable + goal + specifications + content + constraints + output format + ambiguity instructions.

Always review the plan before executing. A 30-second plan-review step prevents 200-credit mistakes.

Pair Manus with other agents for best results. Perplexity for research, Claude Skills for polish, Cowork for filing, Manus for production. Don't try to make Manus do everything.

Three categories of task where Manus is wrong. Quick chat-level questions, predictable repeatable workflows, and high-stakes irreversible work without human review.

The first time Manus delivers a real working artifact — a deployed landing page, a polished pitch deck, a 30-page research report — your relationship with AI changes. You stop seeing AI as "the thing that drafts text" and start seeing it as "the thing that does the work." The cost is real (credits, learning curve, occasional mistakes), but the unlock — getting genuine outputs without paying a freelancer or doing the grunt work yourself — is what the agent era is actually about.

AI Agent (5) — Manus: The Autonomous Agent That Plans and Executes Multi-Step Jobs