Letting Five AI Agents Redesign My Portfolio

The projects section of my portfolio had quietly gotten out of hand. Sixteen projects, each rendered as a big stacked card with a giant decorative number, a sprawling description, and a long row of tags. Individually every card looked fine. Together they formed a wall you had to scroll past with the patience of someone waiting out a commercial break. I knew it needed a rethink, and I also knew I had no taste left for the problem because I had stared at it too long. So I did something a little ridiculous: I asked five different terminal coding agents to redesign it, all at once, and let them argue by example.

The Setup

The five were codex, droid, cline, goose, and kilo. The premise was simple. Give each one the same brief, the same context about the section, and the same goal, a more compact projects layout, and see what each proposed independently. The point was not to crown a winner. The point was to see whether independent agents, with different underlying models and different default instincts, would land in the same place or scatter in five directions.

The one rule I cared about was isolation. Each agent worked in its own scratch directory, a throwaway copy, with no ability to touch my actual files. I was not about to hand five autonomous tools simultaneous write access to the real repository and find out afterward whose change clobbered whose. Isolated directories meant I could read five proposals side by side, calmly, with the real code untouched. If an agent produced nonsense, I deleted the folder and lost nothing.

Two Casualties Before the Bell

Web development being what it is, the experiment did not run clean. Two of the five never really got going. One died on an account-balance error, the provider simply refusing to do more work until someone paid it. The other tripped over a sign-in and auth gap on the GLM-5.1 model, an authentication handshake that never completed, leaving the agent stranded before it could produce anything.

I mention these not to name and shame but because they are the honest texture of orchestrating a panel of agents. You do not get five clean results. You get the ones that worked and the ones that fell over for reasons that have nothing to do with the task, billing and auth and the ordinary friction of stitching together tools from different providers. Planning for attrition is part of the method. If I had bet the whole exercise on a single agent, a coin flip says I draw one of the two that never started.

Three Independent Roads to the Same Place

Here is the part that made the whole detour worthwhile. The three agents that did finish converged. Not loosely, not in spirit. They proposed essentially the same design.

All three split the section into two tiers, a spotlight for a few featured projects shown larger, and a compact index for the rest in a tighter grid. Call it featured-plus-compact-grid, or spotlight plus index. They each independently decided the giant decorative numbers had to go, that they were eating space and adding noise without adding meaning. They each reached for line-clamping the descriptions so every card stayed a predictable height. And they each trimmed the tag lists down to a few signal tags rather than dumping every technology onto every card.

None of these moves is individually surprising. Any of them is something a competent designer might suggest. What is striking is that three separate agents, working in isolation with no knowledge of each other, arrived at the same combination. That is not a coincidence I felt like arguing with.

Convergence Is a Signal

When one agent suggests a design, I weigh it as one opinion, filtered through whatever that model happened to favor that day. Models have defaults, and a single recommendation can easily be a default dressed up as a decision. But when three independent agents, drawing on different models, land on the same structure without coordinating, the suggestion stops reading like one model's bias and starts reading like a property of the problem. The problem has a natural shape, and several different observers found it.

This is the actual lesson, and it generalizes well beyond a portfolio section. Orchestrating a panel of independent agents and watching for agreement is a fundamentally better signal than trusting one model's output and hoping it is right. Disagreement among them tells you where the real judgment calls live. Agreement tells you where the obvious-in-hindsight answer is, the part you can adopt with confidence. A single agent gives you neither signal. It just gives you an answer, and no way to know how load-bearing it is.

What I Actually Shipped

I did not merge any one agent's output wholesale. I synthesized the consensus by hand into a grouped, uniform card layout, taking the shared skeleton the three agents agreed on and fitting it to my own data and visual language. The giant numbers are gone. Descriptions are clamped to a consistent height. Tags are trimmed to what matters. Featured work gets room to breathe up top, and the rest sits in a calm, even grid below instead of a sixteen-card scroll of doom.

The section is shorter, more scannable, and frankly looks like someone made a decision about it, because someone finally did. The agents did not design it for me so much as triangulate the answer I could not see on my own. I still made the final call and wrote the real code. But the panel turned an open-ended taste problem into a question with a clear, well-supported answer, and it cost me an afternoon and a couple of failed sign-ins. I will take that trade again.