The Scam
About a year ago, I handed €7,400 to a web developer to help me build a personal website that was centered around a very specific tool I wanted to develop.
My brief was straightforward: build me a professional personal website to showcase my research, publications, and the new work I was doing at the intersection of AI and psychology. Something clean. Something functional. Somewhere I could finally consolidate years of scattered work and also showcase this new tool we developed to assess the psychological safety of AI systems.
He took the money. Sent a few mockups. And I got excited. Still had to do alot of work to write content and my design specifications etc. But atleast we were on track. But then... He went quiet and eventually disappeared.
After trying various ways and means to get hold of him, I decided to write it off as a loss. But more than the money, I wrote off the idea of having that platform I wanted to help developers to build more psychologically safe AI systems and to empower psychologists to use AI systems in a more ethical manner. But the outlet I needed for my work would have to wait.
The Unexpected Turning Point
Fast forward 11 months to January 2025.
I was building an agentic AI psychological assessment system for a client. A multi-agent orchestrator that could conduct in-depth psychological assessments. Think a multi-layered personality diagnostics and bottom-up wellbeing analysis, the kind that requires coordinating multiple specialized processes running simultaneously.
As I was building and testing this design, Claude 4.6 dropped.
Buried in the release notes was something rather incredible. Anhtropic designed and released the first consumer-available swarm agents. Not the traditional orchestrator model where you manually program each sub-agent and coordinate between them. This was something TOTALLY different. One primary agent could autonomously spawn its own specialized sub-agents that work in collaboration with one another to complete complex tasks.
This was EXACTLY what I needed! It would make the assessment system we were designing so much more intuitive and alot more accurate! But I needed to demonstrate this capability to my client. I needed a real-world, complex project that would showcase what swarm coordination could do.
The problem was that the assessment system I was building didn't have any participant data yet. No interview subjects. And no access to a few hundred people to kinda test and showcase why this approach is a lot better for our use case. So really had nothing to demonstrate the system's capabilities.
And then it hit me: Why not build that website and the tool that I'd given up on last year?
The Build
So I started with how I would build a normal single orchestrator agent. I defined a primary role, created context, gave it a personality and defined its tasks and gave it alot of context. I then defined a few core agents myself. Ones I thought would be the bare basics thats necessary:
- An academic research agent to scan my publications, LinkedIn, entire web presence, and compile a comprehensive picture of my professional identity.
- A personal branding agent to analyze my writing style, values, and positioning across all my articles, white papers, and public communications.
- And a few specialist agents for web scraping, security auditing, image processing, professional writing, and critique.
Then I let the swarm take over. And what happened next just blew me away.
The research agent started its work and started scanning publications, and analyzing my digital footprint. It reviewed my social media and looked over all the articles where I was featured (e.g. Forbes, Psychology Today, Medium etc). But then it paused. It recognized that there was a gap. No matter how much public data it processed, it didn't have the depth it needed to really understand what truly mattered to me.
So it spawned its own sub-agent. Let me say that again, it spawend its own subagent with its own primary task and its own identity. Whilest this primary research agent was still busy searching the web and creating an image of me. The new agent... started to interview me.
I never gave the main research agent that instruction. But for 25 minutes, this sub-agent asked questions I hadn't been asked in years. What's your story? How did you get here? What drives your work? What matters most to you? What are you trying to build? Why is this important to you? Why now?
Man, this didnt feel like a survey. It was a real in-depth interview... about me!
After the interview, the research agent took that interview data and its interpretation of it and combined it with everything the other agent had gathered from the web. It generated a markdown file of just over the 8000 lines!!! It then handed that synthesis over to the branding agent. That agent was already researching UI design, voice, visual story telling etc... But then it analyzed the text. It identified my core values, it determined my tone of voice and writing style, it created its own structure for appropriate visual language, and determined the color palette and typography that would best represent me and my work.
Meanwhile, other agents were already moving:
- One registered a new domain (that was a mistake!).
- Another set up hosting on Vercel.
- Another configured the Neon database.
- Another began migrating content from my business website.
- Another pulled my publication data and citation metrics.
- Another created a script to constantly pull my new publications in from ORCID to my website
- Another found images from my books, videos of my keynote presentations, YouTube videos (some clean, some rough) and processed them for web use.
- Another catalogued every podcast appearance, news feature, and resource I'd ever been mentioned in.
At its peak, I saw 15 agents that were working collaboratively on different aspects of the build at the same time.
Three hours after they started, I had an 85% complete website. Not a prototype. Not a proof of concept. A professional site with automatic publication pulls, integrated blog with Payload CMS, custom AI audit tools, responsive design, and zero breaking elements.
The remaining 15% was refinement, adjusting layouts, perfecting the design language, adding some specialized tools (if you look at my resources page its clear which ones I created and which ones it did :P) and things. I added a fee custom elements, added my personal story and integrated the prototype build of our AI-IARA (Van Zyl, 2026) AI System Audit Tool.
By the end of the weekened, it was live.
Not the End of the Story
But that my friends as not the end of the story. So, I want to say that I was using Claude Code within Cursor to do this. And I stupidly used the "claude --dangerously-skip-permissions" command.
Before I went to bed I asked the one agent to review my site and optimize the meta-data and ensure that everything is aligned to current SEO standards. And.. I went to bed.
The next morning I woke up, made myself a nice cup of coffee and sat down in front of my PC. Completely oblivious as to what chaos occurred last night! On cursor I saw my agent said "task complete".
Cool. Cant remember what I asked, but ok. I opened to log file, took a sip of my coffee and saw it was busy building for 4 hours. First thing that went through my mind was "OMG HOW MANY TOKENS DID I BURN??!?!?!", but then I read what it did and I was amazed.
It pulled four of the psychometric tools we developed, built custom pages for them, designed the assessment forms (fully functional btw!), and then produces some elementary visualizations and scoring instructions. Basically, you can take the assessments and it gave you a very basic output and report.
I asked it, how on earth did it come up with this idea?! And you know what it replied? During the interview I mentioned how frustrating it was to constantly get emails from people asking for permission to use the scales for their research (when it was published open access and under a creative commons license). So when I asked it to make sure that we attended to everything, it went and thought this was outstanding.
So yeah, this was really amazing! Obviously also scary because I didnt put appropriate evals, stop conditions and guardrails in place. But amazed the end product looked so professional.
And do you know how much it cost me to build? 250 euros in tokens!
What This Actually Means
So this isn't a story about saving another €7,200, though I did. And its also not about AI replacing web developers (though it raises questions about what "web development" means when agents can coordinate at this level). This story is about us just crossing a threshold.
We're past the phase where AI can handle discrete, well-defined tasks. We've moved beyond "AI can write your emails" or "AI can generate code snippets." We're entering the era where AI can orchestrate complex, multi-domain projects that previously required coordinating teams of specialists.
The research agent didn't just follow instructions. It recognized what was missing and created the process to fill that gap. The swarm didn't just execute, they adapted, coordinated, and collaborated without constant human intervention.
When an agent can interview you for 25 minutes, synthesize that conversation with years of your public work, hand off insights to a branding specialist, and coordinate with a dozen other agents to build something professional.... we're not talking about an incremental improvement in AI capability. We're talking about fundamentally different possibilities for knowledge work.
The question isn't whether this changes how professional work gets done. I think the question is how quickly you can recognize what just became possible, and what you're going to build with it.
For me, it started with reclaiming a project I'd abandoned. But it ended with understanding that the tools I'm building for psychological assessment aren't just research instruments. They're glimpses of how we'll be working in 2026.
The swarm is here my friends. The only question is what you're going to build with yours.
