Orchestrating agents is a design practice
I shipped a production platform alone — designed, architected, coded. Not because I became an engineer overnight, but because orchestrating AI agents turned out to be something unexpected: a design practice.
What does it feel like to build with an orchestra of agents? Less like programming, more like running a studio. You cast roles, write briefs, direct attention, and review work. The craft is in the direction.
Instruments
The first mistake is treating agents as one instrument. Building Engrym, I ran them as sections of an orchestra: one agent owned data models, one owned interface work, others existed only to attack what the first two produced. Casting is a design act — you are decomposing intent into roles, deciding where judgment must stay human and where it can be delegated. The org chart of your agents is a design artifact, and it deserves the same care as an information architecture.
Briefs
A prompt is a brief. Everything designers learned writing briefs transfers intact: context, constraints, references for taste, and a definition of done. When an agent produces generic work, I re-read the brief before blaming the model — the diagnosis is usually the same as it was with juniors and agencies. The scarce skill is not “prompt engineering.” It is the older discipline of saying, precisely, what good looks like.
A prompt is a brief. Vague briefs produce generic work — from humans and models alike.
Context
Agents fail the way new hires fail: not from lack of talent, but from missing context. They don’t know that the padding is 22 pixels because of a decision made in March, or that the checkout flow is untouchable until the audit closes. I got tired of re-explaining, so I built the explanation into infrastructure: decisions stored once, versioned, and served to every agent at the moment of work. That is information architecture — the right information at the right moment. The consumer changed. The discipline didn’t.
Critique
The loop that makes orchestration work is not generation — it is critique. Nothing ships from a single take. Review agents are briefed to refute: find the flaw, break the claim, fail the build. It is a studio crit, formalized and running at machine speed. Staged disagreement is what turns fluent output into trustworthy output, exactly as it does in a design team.
Taste
What remains scarce when production is abundant? Deciding what should exist, what almost-right means, and what not to ship. Agents raise the floor; taste sets the ceiling. So — is design being automated? The production of artifacts, increasingly, yes. The deciding what should exist and recognizing when it’s right? That just became the whole job.