Introduction
WTF are Agents?
Examples you can try Today
How Agentic Systems Work
The Dark Side: Generative Garbage
Human in the Loop
Remember that scene in The Matrix where Agent Smith could replicate himself and be everywhere at once? Well, we're not quite there yet, Mr. Anderson. But we're entering an era where something similar is happening across all of our computing.
It's 10 PM. You've got a massive project due tomorrow—a research paper, a pitch deck, or code for a new app. You're staring at a blank screen, wondering how you'll pull an all-nighter. Now imagine, a magical fairy shows up to grant your every wish. Type a few sentences explaining what you need, hitting enter, and then... go make a sandwich. When you come back, there's an outline waiting for you. Go make coffee, and return to find research compiled from dozens of sources. By the time you finish your midnight snack, there's a completed first draft with footnotes, references, and even suggestions for revisions.
This is what AI agents can do... right now.
Unlike the chatbots you've been playing with for the past year —where you type something and get a single response back—agents are AI systems that can string together multiple actions, make decisions, and execute complex tasks with minimal supervision.
They're the evolution beyond the one-and-done prompt-response model, and they're about to change everything about how we work.
...
In Hollywood, an agent is someone who represents you, makes deals for you, and then drinks cucumber water while talking into the phone. In tech, an AI agent is somewhat similar—it's a digital entity that represents your interests, but instead of cucumber water, it runs on electricity and a healthy diet of prompts and logically argued guidance.
An agent is essentially a small, standalone program that can operate with some level of autonomy.
It can:
Understand what you want through conversation
Break down complex tasks into steps
Execute those steps by connecting to various tools and databases
Make decisions at critical junctions
Deliver complete outputs instead of just answers
Think of agents as the difference between ordering fast food and having a personal chef.
ChatGPT gives you exactly what you ordered—one burger, coming right up.
vs.
An agent asks what you're in the mood for, checks your dietary restrictions, goes grocery shopping, and then decides to help you with your email, your taxes and your sales pipeline all while planning dinner.
...
What? No-way. Yes, yes-way.
I think the best way to understand something is just to use it.
I recommend doing that, and trying to experiment with something crazy. Make it think, and then reflect on what you just asked it to do. Here's some I've played with.
Manus is an early mover. I improvised something I deliberately thought was really difficult. For me, that was to create a technical paper about "engines and blockchain integration." Instead of spending weeks researching, outlining, and coding, Manus delivered a comprehensive business plan, architecture mock-up, and functional JavaScript code for Unity to web3.js integration in about 25 minutes. That's not just an incremental improvement—it's a paradigm shift.
(They give you enough tokens to try one hard query, the pro plan is 19 bucks.)
Google's Gemini has an agentic mode that can perform multi-step tasks, though I sometimes am weary of Google services. I'm curious how google will evolve in this space especially with their new "Flow" video gen service.
Microsoft's ChatGPT "Vibe Code" capability is evolving in this direction too, likely to integrate smoothly with Office, Visual Studio, and other Microsoft universe of tools. OpenAI just bought the AI IDE Windsurf for "something-Billion," a leader in generative coding. I prototyped with their agentic coder back in December.
And of course, Anthropic's Claude. I have been using Claude now - daily - for almost a year! I am sure with Anthropic's MCP, it's newly created internet search capabilities, and other agentic tools, we will see lots of workflow automations in the Claude ecosystem.
These are but a handful of the agentic workflow services coming out every day.
I've begun to quantify agentic interactions into stages:
Conversational Interface: You start by explaining what you need. The more specific you are, the better, but agents are also getting good at asking clarifying questions.
Understanding & Planning: The agent processes your request and develops a plan. Some will think out loud (ex: Deepseek), breaking down the steps they'll take to accomplish your goal.
Execution: This is where the real work happens. The agent might:
Search the web for relevant information
Read and analyze documentation
Compare approaches from different sources
Write code or text
Test solutions
Create visualizations
Delivery & Refinement: The agent presents its work and can refine it based on your feedback.
I often ask it to comment the code, and provide detailed descriptions for assembly. Essentially, I can have it do the designing research, and allow me to focus on optimizing the pieces and the system.
What makes this revolutionary is that the execution phase can involve dozens or hundreds of discrete actions that previously required human intelligence. An agent researching blockchain technology doesn't just grab the first definition it finds—it might analyze whitepapers, compare transaction speeds across platforms, examine developer communities, check GitHub repositories, and synthesize all this into a cohesive analysis. It actually does the research and learns to refine it's search.
While agentic flows sound complicated to build, the pieces and components are increasingly modular and accessible with non-code tooling. We will potentially see an explosion of automation workflows made in popular software like Langchain, Make and N8N.
There's a flip side to all this.
We're headed toward a world flooded with "high resolution generative garbage."
If you aren't aware, there is a large viral amount of baby podcasts on tik tok. These are entirely generated from AI Agentic Workflows, from script, video, and output.
But automated pipelines generating baby podcasts is only the start - Think fake news articles, automatic youtubes, or just everyone setting their content creation to "auto."
When creating becomes effortless, our digital spaces risk becoming polluted with low-value content masquerading as human-made.
...
This is where you come in.
As designers and creatives, you're at the forefront of defining how these tools will be used.
You can choose to:
Set boundaries: Demand transparency about AI-generated content and advocate for sustainable practices.
Where does the data come from?
Who's compute are we using?
What is the bias in this dataset?
Focus on augmentation, not replacement: Use agents to handle the grunt work while you focus on the truly creative aspects only humans can provide. I maximize editing time, and operational systems. The more control I have on the process the better I can make it.
Shape better tools: Your feedback as early adopters will determine how these systems evolve.
The question isn't whether to use them. We have to. But how to use them in ways that elevate rather than diminish what makes human creativity special. How do make sure everyone benefits from agentic workflows.
So go ahead—try an agent. Approach it as a collaborator. Make your own decisions about it.
The future of work isn't just about what these digital agents can do for us; my hope is it's about what they'll free us to become.
If you vibe to what I talk about, please consider subscribing or sharing with friends.
We'll see you next time.
Nye Warburton is an educator and NooB agentic builder spending his summer playing with AI. This essay was written with improvisational sessions in Otter.ai and then refined and edited with Claude.ai. Images done with Leonardo.ai and Stable Diffusion.
For more information visit https://nyewarburton.com
Here Comes Reinforcement Learning - March 30, 2025
Understanding GPU Tokens - May 4, 2025
Bear_Rabbit_Robot: Mapping AI's Latent Space - March 13, 2025