Per-User Software: The Pipeline’s Death Knell

The one-size-fits-all software distribution model was never a deliberate design choice; it was a cost constraint. Producing correct code was expensive, so teams froze artifacts and shipped one version to everyone. That assumption is collapsing as AI makes code changes cheap, real-time, and even per-user.

Iris ten Teije proposes a stem-plus-divergences model: a canonical codebase with bounded, isolated, and reversible per-user adaptations. This isn't feature flags or segments; it's software that adapts to individual behavior without manual forking. The vision is compelling but introduces hard problems: versioning becomes a graph query, not a number.

Testing, correctness, and desirability become vastly harder when every user runs a unique version. Coordination across millions of divergences requires merging intent, not code. The talk acknowledges these challenges are unsolved—the easy 80% is code generation; the hard 20% is observability, validation, and trust.

For engineering leaders, this represents a paradigm shift in delivery governance. The role of CI/CD pipelines may shrink as the boundary between development and distribution blurs. But without production-grade evidence or benchmarks, the idea remains speculative—a signal worth tracking, not a near-term disruption.