In the Loop: Week Ending 5/10/25
Marketing AI Week in Review – May 4–10, 2025 This week offered a clear view of how quickly AI is moving from experimentation to execution. Big tech an...
A couple weeks ago I wrote about AI agents – what they are and what they aren’t – and shared an experience I had creating an “agent” using Zapier to connect Typeform to HubSpot to Outlook.
As I noted, the agent I created was actually a custom workflow, which is when you train systems to work together to do something for you, with your input and approvals along the way. In other words, these “workflows” are powerful automations, but still rely on constant user guidance.
From Workflows to Autonomy: How Manus Differs from Other AI Tools
An “agentic” system, by contrast, can take the general goal you give it and break the work into smaller steps—often making decisions on the fly. You might have heard of experimental tools like Auto-GPT or BabyAGI, which handle multi-step tasks with minimal human intervention. That’s the key difference: these solutions don’t wait for your sign-off on every little thing.
It’s this promise – that AI systems will be able to act on our behalf – that is driving all the agent hype. And last week I finally experienced it.
A Surprise Invitation: Gaining Early Access to Manus
If you follow AI news, you’ll know that an AI tool called Manus made waves a few weeks ago because it claimed to be the first truly agentic LLM. Buying into the buzz, I added my name to the waitlist, expecting to wait weeks or months for access. So I was surprised when I got the email saying, “You’re in!” less than a week after submitting my request.
First Impressions: Manus vs. Other AI Tools
Much of the Manus interface would be familiar to you if you’ve used the Deep Research tools by Google or ChatGPT. There’s a window where you can tell Manus what you want it to do, a way to upload documents that it can reference, and a list of use cases showing how Manus can be used for different types of tasks.
Once you start a new task, though, you’ll notice something different: the right side of the screen opens up to show you the Manus computers in action, capturing all the activities it’s undertaking to complete your task.
Another difference between Manus and other tools I’ve used is that you’re able to prompt Manus while it’s working to give it new or updated direction. In ChatGPT, for example, once you’ve started a Deep Research project, you have to wait for it to be complete before you can adjust its direction. It’s a small difference, but one that makes the system seem even more like a research partner and not just a task-completer.
Behind the Scenes: Watching Manus Create My Marketing Plan
The biggest difference between Manus and the other tools I’ve tried is its ability to do things entirely on its own that I am fundamentally incapable of.
The first thing I asked Manus to do was create a marketing plan for my company. I gave it parameters, it asked me a bunch of clarifying/focusing questions, then went on its way. I watched as the system created target audience personas (it even gave them names: “Healthcare Hannah” and “Education Edward”), a marketing plan based on my current marketing technology stack, step-by-step instructions on how to implement new workflows in my HubSpot CRM, and other critical elements.
Blown Away by Autonomy: Manus Builds a Website on its Own
While impressive, this kind of functionality exists in other tools I’ve used. What it did next does not: after creating the marketing strategy and implementation plan, Manus asked me if I wanted it to “Turn [my] creation into a website and deploy it permanently.”
I told it yes and watched somewhat slack-jawed as Manus built a website in real time. It created pages, wrote the content, added stock images, created the navigational structure, tested the whole thing – then published it to a temporary server so I could review it. I’m not a developer, so it would have been impossible for me to do this work, but even if I could, I certainly wouldn’t have been able to complete it in less than 10 minutes.
The final question Manus asked me was if I wanted it to publish the website permanently to a live server. I answered “no” to that one, since I wasn’t sure where (or how) it was going to be published, but I can envision a day (soon) where I’ll say “yes” to that question because the AI tool (whether it’s Manus or something else) wants to publish to my own web domain and I’ve given it the access it needs to do so.
Balancing Innovation with Caution: Risks in Autonomous AI
As exciting as it is to hand over these tasks, there are real risks in letting an AI agent act with that level of autonomy. Brand compliance, data security, and legal issues can emerge if you’re giving the system free rein to publish, launch code, or engage with customer data. None of this diminishes the potential – just underscores why we need to balance the convenience of having an agent with the caution of “Is this safe for my brand and my customers?”
Questions About the Future of AI Agents
This experience raises a bunch of questions for me: What is the future of coding? How much control will we trade for efficiency? How much “agency” will we be willing to give AI agents to do things on our behalf? And, finally, with all of this capability at our fingertips, which parts of the work we do today should we keep a firm hand on?
For me, the answer to that last question is pretty simple: I want to maintain control over my business strategy, goals, target audience (Manus can name my personas whatever it would like), and brand (“human in the loop”). But beyond that, I’m happy to have an AI help me with many of the tasks it completed, including creating web pages and testing them for me.
The promise of AI agents isn't fully here (yet), so we have time to ask and answer these questions. But if Manus is any indication, we don't have long.
