Agents move the bottleneck up the stack

InferGo is public and in alpha: a Go-native inference toolkit for backend services. The promise is deliberately narrow: export a supported model once, load it in Go, and run predictions without Python in production.

A lot of the work moved faster with Codex, Claude Code, and Claude Design in the loop. But the speed only held when the contract was clear: bundle versions, parity checks, supported model families, and what should fail loudly instead of pretending to work.

That is the lesson I keep relearning with agents. They reduce typing, searching, and first-pass implementation time. They do not remove engineering judgment. The bottleneck moves up the stack: naming the API, drawing the boundary, choosing the non-goals, and deciding which promises survive the next release.