Mar 12, 2025 · 1h 3m · startup-ideas
Manus AI replaces your AI tech stack? (Full Demo)
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
Host Greg Isenberg and guest Min Choi conduct a live, hands-on demonstration of Manus AI, evaluating its ability to autonomously build software, execute market research, and create browser games. The episode highlights the transformative potential of multi-agent workflows while candidly exploring current deployment bugs, sandbox crashes, and context window limitations.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Greg holds 34.9% of the talking time here. How this is scored →
Greg as informed peer 4.2 Guest teaching 4.8 Guest disagreement 0.1 Greg pushing back 0.6
speaking balance: gold is Greg, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 21:41 Pushback against the 'wrapper' narrative
Min directly dismisses social media claims that Manus is merely an unoriginal wrapper around Claude, arguing it implements an advanced multi-model framework.
Hardest push from Greg ▶ 27:04 Direct inquiry regarding Chinese server securityGreg interrupts the product walkthrough to press Min on data safety concerns and potential government access to sensitive information.
Biggest teaching moment ▶ 44:24 Deep dive into context degradation limitsMin breaks down the exact technical limitations causing the system failure, explaining why long-context agentic operations degrade during deployment.
Greg holds their own ▶ 8:25 Synthesizing the AI cocktail stack workflowGreg demonstrates his domain knowledge by articulating the multi-step developer stack (ChatGPT, v0, Bolt/Replit) and comparing it to Manus's all-in-one execution.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Greg as informed peer | Guest teaching | Guest disagreement | Greg pushing back | Why |
|---|---|---|---|---|---|---|
| Welcoming Min Choi and Exploring Next-Stage Agentic AI | 4 | 5 | 0 | 0 | Greg sets up the episode about Manus AI and invites Min Choi to demonstrate its potential. Min explains the evolution from past agentic AI experiments to the current multi-agent stage, establishing the demo framework. | |
| Setting up the Prompt for a DocuSign Clone | 5 | 3 | 0 | 1 | Greg shares his business premise of disrupting DocuSign with a low-cost AI clone. Min demonstrates how to start simply in the chat interface. | |
| Manus Multi-Agent Architecture and Autonomous Sandbox Execution | 6 | 6 | 0 | 0 | Min explains Manus's internal multi-agent architecture and autonomous sandbox environment, comparing it to BabyAGI. Greg synthesizes the contrast between existing modular workflows and Manus's unified approach. | |
| Refining Project Scope with Real-Time Interventions | 4 | 5 | 0 | 1 | Greg asks whether they can steer Manus mid-execution, prompting Min to show how human-in-the-loop intervention lets them constrain the scope specifically to e-signatures. | |
| Launching a Parallel Session to Mine Greg's Startup Ideas | 5 | 4 | 1 | 1 | While the DocuSign clone runs, Min launches a parallel session to research startup ideas from Greg's content. Greg proposes refining the prompt, but Min prefers letting Manus operate autonomously. | |
| Evaluating Manus AI Architecture, Models, and Non-Technical Accessibility | 4 | 6 | 1 | 0 | Min addresses online claims that Manus is merely a wrapper for Claude, arguing instead that it utilizes an orchestration framework with fine-tuned models and Puppeteer automation to bridge the gap for non-technical users. | |
| Handling Scraping Bottlenecks and Reviewing Code Architecture | 3 | 4 | 0 | 0 | The research session appears stalled on YouTube, prompting a quick status check. Min compares Manus's full project scaffolding to tools like Cursor and Windsurf. | |
| Data Privacy, Security Considerations, and Chinese AI Infrastructure | 4 | 6 | 0 | 2 | Greg directly questions the data privacy risks associated with Chinese servers. Min acknowledges valid data sovereignty concerns, advising users to avoid sharing sensitive financial or personal data while embracing open-source alternatives. | |
| Troubleshooting Sandbox Crashes and Code Completion | 3 | 4 | 0 | 0 | The research sandbox experiences a critical error, requiring a computer reset. Meanwhile, the DocuSign session finishes its coding phase after an ETA prompt. | |
| Navigating Deployment Failures and Self-Healing Debugging | 3 | 5 | 0 | 1 | Public deployment fails in the sandbox, forcing Manus to package the project as a zip file. Min highlights how Manus automatically attempts self-healing debugging during terminal errors. | |
| Contrasting Prompting Methodologies: Concise vs. Detailed Inputs | 4 | 5 | 0 | 0 | Greg observes Min's concise prompting style. Min explains that minimal prompts reduce barrier to entry for general users and lean into agentic inference. | |
| Live SEO Strategy Audit for Late Checkout Agency | 6 | 3 | 0 | 2 | Greg suggests testing Manus on growth marketing for Late Checkout Agency. Greg specifically defines his target audience of enterprise AI executives rather than raw traffic metrics. | |
| Identifying Context Window Limits and Deployment Constraints | 3 | 6 | 0 | 0 | Manus triggers a context window limitation warning, leading to a discussion on technical bottlenecks, long context degradation, and deployment complexities compared to platforms like Replit. | |
| Server Capacity Challenges and Ideal Use Cases for Manus | 4 | 5 | 0 | 0 | Min explains how server overload and viral traffic surges cause frequent timeouts. He categorizes where Manus currently excels (scraping, research, simple apps) versus where it struggles. | |
| Session Limit Workarounds and Building a 3D Flight Game | 3 | 6 | 0 | 0 | After hitting the daily 10-session cap, Min demonstrates a workaround by repurposing a broken existing session to prompt a Three.js flight simulator game. | |
| Live Flight Game Demo, Final Verdict on AGI, and Conclusion | 6 | 4 | 0 | 1 | Greg reviews the post-recording demo of the generated 3D airplane game, assesses the overall state of autonomous agent tools, and closes out the episode. |
Statements from this episode (24)
Isenberg: Creators Can Make $1,000 Daily Building AI DocuSign Clones
“So I think that there's an opportunity to build. I mean, I think multiple people could do this, can make about, you know, at least a thousand dollars a day in revenue.”
Choi: Manus AI operates as a multi-agent system with planners and executors
“Manus is sort of a multi-agent Under the hood. And it's got, like, planner, executor, that type of thing.”
Choi: Manus AI Prompts Users for Help to Bypass CAPTCHAs
“I did, like, run into one interesting thing before was that it got stuck on, like, captchas and stuff, and I thought that was kind of funny, and it wanted me to, like, kind of intervene and, like, do something with it. So it has some sort of the human in the l…”
Choi: Manus AI autonomously browses, researches, plans, and executes beyond coding
“This is a little bit different because it does more than just coding. It actually does the research. It actually goes to the browser, and it actually catches captures the information from there directly, and then it takes all that context into its planning, an…”
Min Choi: Manus AI lets users interrupt and readjust planning mid-execution
“Yeah, this is probably the one of the coolest feature that I've noticed on this particular product is that I can interrupt it in between without it, like, you know, stopping its work that it's going to try to as you see here, it's going to say, I understand yo…”
Choi: Manus AI relies on browser automation tooling like Puppeteer
“It's using a tool called Abiligus Pucketeer which is like an automation tool. So it's, it automates browser tasks. It's a it's a tool that's been around for a while in test automation world so, and it's using that to do a lot of these things.”
Choi: Manus AI is an agent orchestration framework, not a Claude wrapper
“There was a lot of meme going around talking about how, yeah, this is just a wrapper and et cetera. That may be very oversimplifying it. I think what they've done is create like a whole framework under the hood. I believe is that they got a, you know, army of …”
Choi: Manus AI Generates Project Architecture Without Requiring Boilerplate Templates
“Another cool thing about this is that you don't have to have a Like a project structure as you see on other tools, like where, you know, people say, go clone this, you know, template, you know, this boilerplate code, and you can use that to build your app and …”
Choi: Chinese AI Companies Like Manus Must Share Data With Government
“I think the, one of the biggest thing is that if you are a Chinese company I think you're obligated, obligated to give access to the government.”
Choi: Multiple open-source projects are already replicating Manus AI
“And I know there are already a several open source project that popped up that builds around what Manus has done.”
Min Choi: Manus AI retains progress across sandbox resets
“The cool thing is that it remembers where it left off.”
Choi: Manus AI deploys apps directly to subdomains on its domain
“What that's gonna do is, it's got its own public, like, similar to what, like, Bercel rep lit has it's got its own little deployable website, and it's gonna give you a subdomain with their primary domain attached to it so you can actually see the website in ac…”
Choi: Most AI users prefer abstracted black-box systems over manual controls
“It's a very abstracted very, you know, black box. Some may not, some may like that and some may not like that. But I think for the most users, you know, they'd rather have it, you know, black box. They don't want to worry and think about it.”
Choi: Long-context degradation is Manus AI's biggest pre-launch issue
“And I think that this is probably the biggest issue that Manus AI team needs to probably address before it goes it goes live. Cause I think that a lot of people are going to run into this similar issues, especially power user like me, I'm gonna, you know, keep…”
Min Choi: Replit executes AI deployment well and its agent is improving
“Deployment is, is not as straightforward with these tools and I know Repl.it has done that really well and I think that's the, it's got some good future in, you know, where, the direction that it's trying to go their AI agent is, is, is getting better and bett…”
Min Choi: Manus AI is trying to do too much and hasn't solved deployment
“Manus has to work on their deployment process as well. So I think they're just trying to do too much. I think that's part of the problem is they're trying to have all of this kind of package it, like all this services and products are kind of packaged into thi…”
Choi: Manus AI Went Viral in the US Two Days Ago
“They actually launched like six days ago, but it didn't like go viral until like two days ago, basically. So when it came over to U S and people in the U S start to get access to it, and that's when it started to get like real traction.”
Choi: Manus AI excels at research, simple games, and landing pages
“The research one works really well, obviously. You know, you expect that to work well. Simple games work well landing pages work well. And I think that's sort of the ones that have been very successful with Manus AI so far.”
Choi: Manus AI struggles with complex applications like DocuSign
“The, as you get more complex like DocuSign and stuff like it does have its challenges”
Choi: AI-Generated Code Should Only Run in Sandboxes, Not Locally
“I don't download this stuff into my local computer. I have a, I just create like a virtual sandbox on the cloud and then I run it there just in case as well.”
Choi: Manus AI restricts beta users to 10 requests or sessions daily
“There is a, I believe there's a 10 request or 10 sessions limit in the day.”
Choi: Users Can Bypass Manus AI Session Limits Using Open Threads
“When you reach a maximum limit, just a different session and it will, you can tell to do something else.”