Summary
Iris ten Teije, co-founder of Differ and formerly an early engineer at a fintech bank that scaled and exited, argues at the AI Engineer conference that the traditional one-way software pipeline — build a frozen artifact, ship one version to everyone — is not a law of software but a consequence of economics that has now changed. When producing a correct code change cost skilled engineers days, shipping a frozen artifact to all users was the only sensible architecture. As AI collapses that production cost toward zero, the underlying assumption dissolves.
Differ’s proposed alternative: rather than a single monolithic artifact controlled by feature flags, companies deploy a canonical ‘stem’ and each user runs their own bounded, isolated, individually reversible divergence of it. Adaptive software can be modified in real time within developer-defined boundaries — specific form fields or payment flows can be marked off-limits while other elements adapt to user context.
Ten Teije directly addresses the obvious objection — that AI-generated codebases are already hard to reason about — by distinguishing unmanaged divergence inside a tangled artifact (brittle) from isolated divergences with a blast radius of one user session (safe and reversible without a deploy). She cites forward-deployed Salesforce engineers and enterprise professional services as evidence that per-user software customization has always been in demand; only the economics prevented it from scaling.
📺 Source: AI Engineer · Published July 07, 2026
🏷️ Format: Opinion Editorial
