Better-Trained Limbs
A year and a half of drums and agent systems on the same learning curve. The goal turned out to be better-trained limbs, not a smarter orchestrator.
I’ve been learning drums for a year and a half. I’ve only been building agent systems since March. The learning curve is identical.
I realized this in the shower this morning. Bear with me.
Stage 1: You’re coordinating, and it’s hard
When you first sit behind a kit, your brain is doing everything manually. Right hand on the hi-hat. Left hand on the snare. Right foot on the kick. Left foot on the hi-hat pedal. Every hit is a conscious decision. Nothing flows.
If you try to go faster than you can handle, your limbs start lurching on their own. Your foot kicks when it shouldn’t. Your hand rushes the beat. The harder you push, the worse it gets. You have to slow down and build the muscle memory before you can speed back up.
Early agent systems work the same way. It’s easy to get excited and start building something complex before the basics are solid. The system starts misfiring. Steps run out of order, loop, or fail quietly. The fix is the same: slow down, get the basics right, then build up.
Stage 2: Simple grooves, in time
After enough slow reps, something clicks. You can play a basic groove and actually keep time. It’s not impressive but it’s stable, and stable is the whole point at this stage. Your brain stops managing every individual hit and starts holding the pattern.
Early agent builds feel the same. A simple workflow that runs reliably end to end. Not complex, but consistent. The orchestrator is still doing most of the work but the thing actually works, which is a bigger deal than it sounds.
Stage 3: Technique develops
This one’s less obvious. There’s a difference between hitting the drum and knowing how to hit it. Real technique means working with the stick’s natural momentum instead of fighting it. The drum head wants to push the stick back up. Let it. You stop forcing every note and the whole thing gets faster, cleaner, and a lot less exhausting.
Same trap in agent systems. Early on you’re forcing everything: hardcoding flow control, catching errors the model could fix on its own, validating outputs the model wouldn’t have gotten wrong in the first place. The technique is learning what to stop doing. Where to let the system bounce back instead of holding it down. Less fighting, more flow.
Stage 4: Dynamics
Now the limbs are coordinated and the technique is there. The next layer is dynamics. Ghost notes instead of full snare hits. Rim shots where you need the crack. Soft where the song breathes, loud where it needs to punch. Playing the right notes at the right time is table stakes. Playing them with the right feel is what makes it sound like music.
This is the layer most agent builds skip. A system can be technically correct and still feel robotic and brittle. Knowing when to escalate to a bigger model, when to hand a thing to a human, when a short blunt answer beats a polished one. That’s the difference between a system that runs and a system that anyone would use twice.
Stage 5: From solo to ensemble
Polyrhythm is each limb playing its own independent rhythm at the same time. Your brain is receiving different signals from three or four things at once, each doing something different, and somehow holding it all together. I can barely play two conflicting rhythms. Five feels like a distant fantasy.
This is where the analogy starts to break, and I should say so. Agent systems do something that looks similar but is actually easier in a specific way. You’re not one player playing five rhythms. You’re conducting a band, where each agent already knows its own pattern. The work shifts from doing the playing to keeping everyone in time. Different problem, similar feel from the seat.
The drummer as gatekeeper
In a band, the drummer is the timekeeper. They play and they control the pace of everything that reaches the audience. Too fast and the whole thing falls apart. Too slow and it drags.
LLMOps, too. Something has to govern throughput, rate limits, and what actually makes it out to the end user. In my own systems, that’s usually the orchestrator and a token budget acting together. The drummer doesn’t decide what the song says, but they shape how it lands.
I don’t know if this analogy holds at every seam. But it has me thinking about agent design differently. The goal is probably better-trained limbs, not a smarter orchestrator.