I spent a few weeks building my own agent orchestrator. I called it Nerve — a dispatcher that hands work to a fleet of workers and manages its own memory, wake-ups, and state. It worked. Then Anthropic shipped dynamic workflows in Claude. A few weeks later, the platform had built most of the orchestration layer I’d just hand-wired.

This isn’t a one-off. It’s a pattern.

Let me say what a harness is first. The model is the engine; the harness is the scaffolding you bolt around it: orchestration (who runs first, who runs next), memory (not going amnesiac across sessions), scheduling (waking up at the right time to do the right thing). The model doesn’t come with any of this, so everyone building agents this past year has been hand-building a harness. I did too.

But a harness has a fate baked in: it patches what the platform is missing, and the platform always absorbs those patches. The problems I solved building Nerve — parallel dispatch, sub-task orchestration, structured output — were all, in hindsight, general problems the platform was bound to solve. I was just a few weeks ahead of the official version. A few weeks isn’t a moat. It’s a lead time.

So the real question was never “can I build this.” It’s “will the official side build this into the next version.”

The things that will get absorbed — generic orchestration, generic memory, generic scheduling — don’t hand-build them. You’re racing the platform roadmap, you’re going to get absorbed, just wait.

The things that won’t — your private data, your specific business workflow, your distribution channels, your judgment and taste — the official side will never know these. That’s where you build.

I’m not saying Nerve was wasted. Building it is how I learned which lines get absorbed and which don’t, and that read alone paid for the ticket. But if you’re still spending weeks hand-building a generic orchestrator expecting it to be a wall — it isn’t. It’s the next release note.

To judge whether an agent project is worth building, ask one question first — “will the official side build this into the next version?” If yes, don’t build it, wait. If no — because it depends on context only you have — that’s your moat. The harness era is ending, and it’s time to build where the model can’t reach.