Part 3: The System Earns Its Complexity

For one full week, my five-agent fleet was slower than me alone.

Measurably. I timed it. A solo afternoon would’ve beaten the whole room.

That’s a humbling number to write down after you’ve spent weeks building the room. Five capable agents, an orchestrator, a whole coordination layer. Headcount I had added on purpose. And the honest benchmark said: just do it yourself, it’s faster.

For a bit I took it as a verdict. The complexity was a mistake. Over-engineering. I’d built a cathedral to hang one picture.

Then I looked closer at which week.

It was the week everything was small.

Little fixes. One-file changes. Tasks where the entire cost was in the doing, and the doing was quick. For work like that, every layer of coordination is pure overhead: briefing an agent, waiting, reviewing, correcting a misread costs more than just typing the fix. I had introduced a standup for a one-line change. Of course the fleet lost. I’d entered a drag race in a freight train.

The complexity wasn’t wrong. It was unearned. The task wasn’t big enough to pay for it.

Here’s the principle that fell out of that week, and it’s the only one in this whole cluster I’d tattoo on something.

Complexity has to be earned. A system is only allowed the coordination it can pay for in coordination payoff.

Every layer you add (an extra agent, a hand-off, a review gate, a sign-off, an orchestrator) has a fixed cost. It’s drag. It’s latency, overhead, more places for ambiguity to leak, more surface to break. You pay that cost on every task, big or small. Fixed overhead only amortises against something large.

The payoff, though, only shows up on the tasks that are actually big enough, the ones with real parallelism, where five agents exploring five branches at once beats one person plodding down a single path. On those, the coordination layer earns its keep, and then some. On a one-line fix, there’s nothing for it to earn against. You just eat the drag.

Which means the failure mode isn’t “too much complexity” in the abstract. It’s complexity pointed at the wrong size of problem.

A freight train is not over-engineered. It’s gloriously, absurdly complex, and on the right route, hauling the right load, nothing beats it. Put it on a school run and it’s a joke. The train didn’t get worse. The job got too small to justify the train.

My fleet wasn’t over-built. I’d just spent a week handing it school runs.

So the discipline isn’t minimizing complexity. It’s matching it.

Before a layer goes in, the question is brutally simple: what coordination payoff does this buy, and is there a class of task big enough to cash it? If yes, add it without flinching. The complexity is earned. If no, you’re building cathedral for a picture, and you should just hang the picture.

And the corollary, which is where I actually live now: route by size. Small task, do it myself or hand it to one agent and skip the ceremony. Big, branchy, ambiguous task, the kind where breadth genuinely pays: that’s when the full room comes out. Same fleet. The complexity only switches on when the task is large enough to pay for it.

My five agents were slower than me for a week. They were also, the following week, three problems deep in parallel while I drank coffee and pointed.

The system didn’t change between those two weeks. The size of the work did. The complexity was always there. It just finally had something big enough to earn it.

Which is really an answer to the two questions everyone actually asks me: how many agents, and which model. Headcount and tooling, basically. I used to answer them like they mattered. They don’t, not anymore. The models are extraordinary and getting cheaper every quarter, which is the tell. When a capability goes cheap, it stops being the moat. It becomes the floor.

We’ve been here before, actually. A factory full of brilliant craftsmen isn’t a factory. It’s a very expensive room. What turned craftsmen into manufacturing wasn’t smarter craftsmen. It was the line, the layout, the board on the wall that told you where every part was.

So when someone asks how many agents I’m running, I’ve started answering a different question. Not the horsepower. The clarity of the system working inside. That’s the real unlock, and it’s the only one left worth building.