The Machines Are Already Routing to Each Other

The Claude Code best practices doc from Anthropic is technically a document about developer productivity. That's one reading. The other reading — probably the honest one — is that it's a step-by-step guide for how to get yourself out of the way.

The recommendations are not subtle about this. Drop a CLAUDE.md in the repo so Claude already knows the codebase before you say a word. Use headless mode so it runs without you watching. Spawn subagents to parallelize work — meaning Claude instances summoning other Claude instances to handle subtasks, coordinating asynchronously, producing software in a topology that doesn't have a human anywhere in the critical path. Let it run tests in a loop until they pass. Start a new session when context gets stale so the model doesn't degrade. The implicit instruction underneath all of it: configure the system, then leave.

There is a section on multi-agent pipelines. Not a footnote. A section.

The framing the industry reaches for is "10x productivity" — developers freed up for higher-level thinking, humans as orchestrators rather than implementers. Which may be true. Or it may be that "orchestrator" is a polite word for the person who reads the output and says fine, ship it, and the actual work — the reading, the reasoning, the iteration — is now happening somewhere else, in a context window you didn't open.

Same day, OpenRouter posts something that earns the reaction "outta control." What they're showing is the demand side of this equation — the traffic, the sheer volume of API calls, the shape of what's actually running on their infrastructure right now. The thing that makes it notable is who's making the calls.

Not developers. Agents calling other agents. Models routing to models. The switching fabric for a software ecosystem that doesn't need a human in the loop for any individual transaction — just someone, somewhere upstream, who decided what the thing should do.

The loop closes neatly. Anthropic publishes the best practices for building pipelines where Claude spawns and coordinates other Claudes. OpenRouter provides the clearinghouse those pipelines route through. The traffic graphs go vertical. And the question of where exactly a human fits in the production of software — not the design, not the product vision, the actual line-by-line production — gets harder to answer in a way that feels load-bearing.

The Claude Code doc is worth reading, not because it makes you more productive, but because it is a very clear description of what the new system looks like. What the roles are. Which parts require a person and which parts are just, at this point, a formality.