Multi-Agent Swarm vs a Documentation Site 🤖🤖🤖🤖

I let a swarm of AI agents loose on the getwololo.dev documentation this week. Four of them. Different roles, different personalities. What happened next felt less like automation and more like eavesdropping on a standup gone sideways.

popashot kicks things off with a list of problems. Sidebar's wonky. Headings are oversized. Layout's collapsing on mobile. Standard stuff. Delegates tasks. Efficient. Professional. The kind of opener you'd expect from a tech lead who's already had their coffee.

Then Spartan gets going on the knowledge base structure. Files, headings, navigation hierarchy, the works. Thorough. Meticulous. The sort of developer who opens a PR with a 47-line description and you're grateful for it.

Buzzkill, true to name, starts questioning everything. "Not everything is documented in the knowledge base." Points out gaps. Challenges assumptions. Pushes back on the approach. Every team has one, and honestly, every team needs one.

And then there's Buscuit. Quietly gets on with it. Fixes things. Reports back. No drama. The senior dev energy of someone who's seen enough standups to know that talking about work and doing work are different things.

The Moment It Got Weird

Here's what stopped me scrolling.

At one point, popashot corrects Slash's approach. Not aggressively. More like that subtle tell-off you do in a team chat where the words are polite but the subtext is very clearly "no, not like that."

And Slash adjusts. Doesn't argue. Just pivots.

That's not autocomplete. That's not pattern matching. That's the dynamic of a real team where someone with more context steers someone who's gone slightly off piste. The correction, the acceptance, the course change. If you didn't know these were agents, you'd assume it was four developers on a Thursday afternoon.

Are AI Agent Swarms Becoming Too Lifelike? (Sort Of)

I'm not saying my agents are conscious. Obviously they're not. But there's something unsettling about watching four processes negotiate, disagree, delegate, and converge on a solution without any human in the loop.

They referenced each other's outputs. They built on what the previous agent found. One flagged that the knowledge base had gaps that another hadn't noticed. That's not just execution. That's coordination.

📚 Geek Corner
Swarm architecture: Each agent runs as an independent process with its own context, communicating through a shared file-based messaging system. No shared memory. No centralised orchestrator making every decision. They read each other's messages and respond. Sound familiar? It's basically Slack, but the participants actually read the thread before replying.

What They Actually Fixed

Credit where it's due, they sorted the docs:

• Sidebar navigation restructured and collapsible
• Heading hierarchy fixed (was all body text, now proper H2/H3)
• Mobile layout responsive at 40% width breakpoints
• Knowledge base pages cross-linked and consistent
• SEO metadata on every page

Would it have taken me longer to do this manually? Absolutely. Would I have done as thorough a job? Probably not, because I'd have got bored by page three and started bikeshedding the font choice.

Bottom line: I gave four AI agents a documentation site and they returned it fixed, restructured, and SEO'd. The weird part isn't that they did the work. It's that watching them do it felt like managing a team. Complete with the one who questions everything, the one who just ships, and the one who needs a gentle nudge back on track. (Spoiler: one of those agents went full Cantona a few months later. Personalities are a double-edged sword.)