Build a student survey form with Manus
Having just got my hands on the new general AI agent from Manus I was excited to see what it could do. Task: Having just got my hands on the new general AI agent from Manus I was excited to see what it could do. So I asked it to design, conduct and write up a human subjects research study – with no humans! Top takeaways (generated by Perplexity) Everything you’re about to read is made up. Sort of. I’ve been playing with Manus this past week — a new agentic AI from the Chinese company Monica, and a product the company’s co-founder and chief scientist Yichai “Peak” Ji describes as the “the first general AI agent.” Ji may be over hyping the “first” a bit. But Manus is, nevertheless quite startling in how it’s able to devise and execute a plan based on an initial prompt. There are plenty of examples of what Manus can do on the product’s website: planning travel, creating courses, researching products — the usual examples of what a web-based agentic AI can do in other words. But I must confess that I’m looking for a lot more from a “general agentic AI.” And so I decided, rather boldly, to ask it to conceive, design, and conduct, a complete research study! The aim here was to see how capable Manus is in interpreting instructions, setting its own goals, and executing a plan. The actual research study was simply the mechanism for doing this — and this is where the “made up stuff” comes in. That said, the “made up stuff” was pretty good — as was the ability of Manus to deliver on what I asked of it. Manus hit the scene a couple of weeks ago with a select number of people getting early access, and has been causing quite a stir since. Rather than a single AI, it’s a collection of AI tools that are overseen by a central AI “task master” which translates requests into tasks that are then executed by different agents — all while autonomously reviewing and revising the plan along the way. All of this runs on a virtual computer which is able to autonomously open and interact with web pages, allowing Manus to use the web to research, plan, execute, and report back on the things it’s asked to do. If this sounds a little like what companies like OpenAI, Anthropic and others are working on, it’s because it is. Except that Manus’s approach of using a hierarchy of multiple AIs and its own virtual computer feels substantially different to the competition And this came through in spades as I set it the task of conducting its own research study. A human subjects study without human subjects For my test case, I set out to see how Manus would handle with a request to design and conduct a survey-based study on attitudes toward artificial general intelligence (AGI). The task is interesting as it depends on the AI’s ability to interpret a somewhat vague request, craft a solid survey tool, analyze collected data, and write up the results. Running such a study with real human subjects would take a lot of time and effort. Apart from the Institutional Review Board hoops this would entail, it would require funding, time (lots of it), and far more patience than I currently have to find people willing to participate. So I also asked Manus to create a bunch of synthetic students to run the study with! At this point it should be clear where the “made up” part of this post begins to kick in. But even though all of the data used were fake, there was method in my madness. From the outset, I was interested in four things: Could Manus create, from scratch, a survey tool in Google Forms designed to research undergraduate student attitudes to AGI? Could Manus create, from scratch, a survey tool in Google Forms designed to research undergraduate student attitudes to AGI? Could it simulate 1,000 students completing the survey — essentially creating a synthetic dataset and using this in the study? Could it simulate 1,000 students completing the survey — essentially creating a synthetic dataset and using this in the study? Could it carry out a comprehensive analysis of the resulting data and provide a polished final report? And Could it carry out a comprehensive analysis of the resulting data and provide a polished final report? And Could it do all of this from a very simple initial request? Could it do all of this from a very simple initial request?; tools: Manus. Human role: The original post does not say how a person stayed involved.; limitations: Submitting synthetic data to real forms triggers Google's anti-spam measures and violates research ethics. Manus operates within a virtual browser environment, limiting direct API interactions. The agent may autonomously refuse tasks it deems unethical, requiring human redirection. Results are not valid for real-world decision-making..
Tools
Published on Agent Work Atlas:
Overview
Having just got my hands on the new general AI agent from Manus I was excited to see what it could do.
Workflow
Having just got my hands on the new general AI agent from Manus I was excited to see what it could do. So I asked it to design, conduct and write up a human subjects research study – with no humans! Top takeaways (generated by Perplexity) Everything you’re about to read is made up. Sort of. I’ve been playing with Manus this past week — a new agentic AI from the Chinese company Monica, and a product the company’s co-founder and chief scientist Yichai “Peak” Ji describes as the “the first general AI agent.” Ji may be over hyping the “first” a bit. But Manus is, nevertheless quite startling in how it’s able to devise and execute a plan based on an initial prompt. There are plenty of examples of what Manus can do on the product’s website: planning travel, creating courses, researching products — the usual examples of what a web-based agentic AI can do in other words. But I must confess that I’m looking for a lot more from a “general agentic AI.” And so I decided, rather boldly, to ask it to conceive, design, and conduct, a complete research study! The aim here was to see how capable Manus is in interpreting instructions, setting its own goals, and executing a plan. The actual research study was simply the mechanism for doing this — and this is where the “made up stuff” comes in. That said, the “made up stuff” was pretty good — as was the ability of Manus to deliver on what I asked of it. Manus hit the scene a couple of weeks ago with a select number of people getting early access, and has been causing quite a stir since. Rather than a single AI, it’s a collection of AI tools that are overseen by a central AI “task master” which translates requests into tasks that are then executed by different agents — all while autonomously reviewing and revising the plan along the way. All of this runs on a virtual computer which is able to autonomously open and interact with web pages, allowing Manus to use the web to research, plan, execute, and report back on the things it’s asked to do. If this sounds a little like what companies like OpenAI, Anthropic and others are working on, it’s because it is. Except that Manus’s approach of using a hierarchy of multiple AIs and its own virtual computer feels substantially different to the competition And this came through in spades as I set it the task of conducting its own research study. A human subjects study without human subjects For my test case, I set out to see how Manus would handle with a request to design and conduct a survey-based study on attitudes toward artificial general intelligence (AGI). The task is interesting as it depends on the AI’s ability to interpret a somewhat vague request, craft a solid survey tool, analyze collected data, and write up the results. Running such a study with real human subjects would take a lot of time and effort. Apart from the Institutional Review Board hoops this would entail, it would require funding, time (lots of it), and far more patience than I currently have to find people willing to participate. So I also asked Manus to create a bunch of synthetic students to run the study with! At this point it should be clear where the “made up” part of this post begins to kick in. But even though all of the data used were fake, there was method in my madness. From the outset, I was interested in four things: Could Manus create, from scratch, a survey tool in Google Forms designed to research undergraduate student attitudes to AGI? Could Manus create, from scratch, a survey tool in Google Forms designed to research undergraduate student attitudes to AGI? Could it simulate 1,000 students completing the survey — essentially creating a synthetic dataset and using this in the study? Could it simulate 1,000 students completing the survey — essentially creating a synthetic dataset and using this in the study? Could it carry out a comprehensive analysis of the resulting data and provide a polished final report? And Could it carry out a comprehensive analysis of the resulting data and provide a polished final report? And Could it do all of this from a very simple initial request? Could it do all of this from a very simple initial request?
Reported result
Manus produced a 44-page report based on simulated data, including metrics like 53.4% expecting AGI within 25 years. The agent successfully adapted its plan when blocked from submitting data to Google Forms. The author found the report quality impressive but emphasized that all findings are fabricated and serve only as a capability demonstration.
Human role
The original post does not say how a person stayed involved.
Limitations
Submitting synthetic data to real forms triggers Google's anti-spam measures and violates research ethics. Manus operates within a virtual browser environment, limiting direct API interactions. The agent may autonomously refuse tasks it deems unethical, requiring human redirection. Results are not valid for real-world decision-making.
Try it
Act as a research assistant using Manus. I want to test your ability to design a survey and analyze synthetic data. First, create a Google Apps Script to generate a survey form about [replaceable topic] for [replaceable target audience]. Do not submit any data to the live form. Instead, generate a synthetic dataset of [replaceable number] responses within your virtual environment. Analyze this data to identify key trends and correlations. Produce a draft report summarizing the findings. Before finalizing, show me the script and a sample of the synthetic data for my approval. If you encounter ethical blocks on data submission, pivot to internal analysis immediately. Ensure the report clearly labels all data as simulated.
Source reports are not independently verified. Publication date is the catalog publication date, not the original article date.