An orchestrator layer is the routing tier that decides which of an AI agent’s skills load for a task, in what order, and which gates can block the result — before any of the work starts. Mine sits on top of the 80-skill registry I run Claude with every day. It is the least glamorous thing I have built this year, and the highest-leverage.
What is an orchestrator layer?
A skill, in Claude’s world, is a reusable instruction file: a short description the model always sees, and a body that loads only when the skill fires. That split is the entire economics of the system — descriptions are paid for on every turn, bodies are free until invoked. An orchestrator layer is the thing that reads an incoming task, classifies the problem, and routes it: which skills load, what order they chain in, and which blocking checks stand between “done” and “shipped”. Without one, skill selection is vibes, and vibes stop scaling at about a dozen skills.
Why does one agent need 80 skills?
It doesn’t need 80. It needs the right one, instantly. The registry peaked at 107 skills; a consolidation pass on August 1 folded it to 80 by absorbing narrow skills into broader owners. That fold produced the sharpest rule in the whole system: when skill B absorbs skill A, B’s description has to carry A’s trigger phrases, or the capability becomes unreachable — still on disk, invisible to routing. A merge that loses the trigger words is a silent amputation.
I have written before about designing around a constraint until the constraint becomes the design — when the rate limit is the product — and the registry is the same shape. Descriptions cost tokens every single turn. That one fact forces everything: fewer skills, denser descriptions, and an orchestrator to spend the budget well.
How does the routing actually work?
Two tiers fire without being asked. Gates are blocking checks that run after an action: a pre-ship ritual before anything is pushed or uploaded, a review gate that stress-tests a change before it counts as done. Always-on skills are standing disciplines: a problem classifier, a context distiller that keeps token spend flat, a surgical-edit rule that forbids blind rewrites. Everything else — most of the registry — loads only when the orchestrator chains it in, and a multi-phase chain writes its state to a shared blackboard file so the task survives its own length. The orchestrator’s real job is refusal: deciding what does not load.
What broke without it
Every rule above is a scar. The same async re-entrancy bug shipped twice before it became a skill instead of a memory. A hand-maintained index claimed 19 entries while 24 existed, which taught me that a count you produce is a claim, not evidence — verify by a second method or you have verified nothing. One early fold quietly stripped a feature because the absorbing skill’s description never mentioned it. None of these were fixed by trying harder. They were fixed by turning the failure into infrastructure: a gate, a ledger row, a routing rule. The longer version of that philosophy is in How I Build.
Where this is going
Everything public on this site — the games, the extensions, the tooling — is built with Claude, and the direction of travel seems obvious from where I sit: the model stopped being my bottleneck a while ago; orchestration is. The interesting engineering has moved one layer up — routing, gates, and state. If you are building with an agent and not building the layer that tells it what to load, then you are the orchestrator layer, and you don’t scale either.
Three Claude guest passes, free
I have three Claude guest passes to give away. Each one gives a friend a free week of Claude Cowork — including Claude in Excel, PowerPoint, Chrome, and Claude Code. Want one? Leave a comment describing the first thing you would orchestrate, or share this post somewhere I can see it. I will send pass links to the three answers that teach me something.
Fine print, all of it honest: there are exactly three; selection is my judgment; the passes are referral links, so if a winner later subscribes I receive $10 in usage credits; and this giveaway is mine, not Anthropic’s.
The full devlog archive is on the history page, and shorter Q&A lives in answers. Comments are open.
Leave a Reply