‘Deploying AI agents’ sounds so hi tech and futuristic to (non Comp-Sci) me whilst weirdly also resonating of classic 60s and 70s TV shows I loved as a kid. I have been fiddling for a while on the blurred boundaries between LLMs and Agents, notably with Claude, but what appealed when I first saw Manus was the execution of outputs seemingly beyond what Claude can manage. Funnily enough it looks quite a bit like Claude but it seems it is actually a multi-tool agent. I pretty much concur with the conclusion from the MIT Tech review:

While it occasionally lacks understanding of what it’s being asked to do, makes incorrect assumptions, or cuts corners to expedite tasks, it explains its reasoning clearly, is remarkably adaptable, and can improve substantially when provided with detailed instructions or feedback. Ultimately, it’s promising but not perfect.

Caiwei Chen

Anyway, I finally got in, having been on the Manus waitlist for a while. Developed by Chinese startup Monica, it is an autonomous AI agent capable of executing complex online tasks without ongoing human input and created something of a buzz. TL:DR: This is the initial output from first prompt to web-based execution. The selection and categorisation need honing but this in my view is an impressive output. The second version after addition of a follow up prompt.

Longer version:

I wanted to see what I could get from a single prompt so decided to see if it could build a shareable, searchable web page that curates short how-to videos (under five minutes) by higher education educators demonstrating uses of Generative AI. I began by requesting Manus to collect and cluster videos showing how AI is applied in teaching, assessment, feedback, and research (Natural Language Prompt). Manus responded immediately by creating a structured project directory and initiating web searches to identify relevant video content, starting with collections from institutions like Notre Dame and Harvard (which it didn’t get beyond in the first iteration).

Once videos were sourced, Manus automatically filtered them to ensure they were under five minutes in length (but failed to note that one wasn’t a video and others linked to same page where they were already curated!) and produced by educators in the higher education sector (this it did get right). It then categorised them by thematic area (Teaching, Assessment & Feedback, AI Literacy, and Research but the categrosations were not always spot on) while also tagging institutional affiliations (mostly OK). The tagging and filtering work pretty well but the actual sourcing is very limited in that first iteration. A metadata database was created to support an interactive interface, allowing users to search videos by presenter, title, or description and filter by category or institution. Manus built this interface and deployed the collection to a live web page: https://cjubrvkx.manus.space. It took about 15 minutes. Iteration is by natural language prompting for fine tuning.

After launch, I expanded the scope by asking Manus to include contributions from additional individuals and institutions- I threw in my old buddy Dr. Gerhard Kristandl (Dr GeeKay) who I know has a YouTube channel, I added me- Martin Compton (without any additional info) and it found relevant content from each of us. The other three potential souces I threw in reaped no reward though. This is the reasoning:

I searched for content from Tadhg Blommerde, University of Sydney, and Monash University as well, but most of their videos exceeded our 5-minute duration criteria. I’ve documented all findings in the data directory for reference.

Manus resumed its asynchronous task execution, searching for and integrating further video content into the existing framework enabling me to publish a second version . So the output works, I would need much more thoughtful prompting to get it how I want it and, tbh, there are better ways of curating YouTube content but the principle is what interested me and the output is remarkable given the stages and processes it went through.

You can watch a replay of the agent in action here. Not as much fun as watching Napoleon Solo and Illya Kuryakin combat THRUSH (I know, I know).