Iris ten Teije warns one-size-fits-all software distribution is dying
Why the standard CI pipeline is losing its purpose
For decades, we treated the build-and-deploy pipeline as an unshakeable law of physics. We write code, package it into a container image, push it through a CI/CD pipeline, and ship that frozen artifact to everyone. Iris ten Teije, co-founder of Differ, points out that this setup only exists because producing a code change was historically slow, risky, and expensive. When changes cost thousands of dollars in human labor, you freeze the artifact to guarantee safety.
Today, that economic reality is collapsing. When code generation drops to near-zero cost, the barrier between writing software and running it vanishes. We no longer need to ship the exact same binary to every user.
The stem and divergence architecture
Instead of gating a massive, single codebase with feature flags, a new design pattern is emerging: stem plus divergence. In this architecture, a team deploys one canonical "stem" version of the application. As individual users interact with the app, the system observes their behavior and spawns localized, isolated "divergences" tailored to their needs.
Take a CRM tool built originally for sales teams. If an investor uses it, they might constantly log intro paths while ignoring standard sales pipeline fields. An adaptive system recognizes this pattern, hides the unused fields, and surfaces the relevant paths. This is not simple configuration. The runtime environment actually modifies the running code.
Taming the chaos of millions of versions
Running unique software versions for millions of users sounds like a maintenance nightmare. A skeptical CTO might ask how anyone can debug a system when no two users run the same code.
The solution lies in strict architectural boundaries. Because these divergences are isolated and bounded, they cannot corrupt the canonical stem or leak into other users' environments. The blast radius of any automated code change is exactly one user session. If a divergence breaks, the system rolls it back instantly without requiring a full deployment.
The real engineering hurdles
Generating code is the easy part. The actual work of adaptive software lies in solving the infrastructure challenges:
Redefining the source of truth
When version numbers no longer exist, checking what a user is running becomes a graph query. Every divergence must be immutable, attributable, and traceable back to the specific signal that triggered it.
Merging intent over code
When the engineering team updates the main stem, those changes must propagate to active divergences. The system cannot merge raw code lines. Instead, it must merge intent and outcome, ensuring everyone converges on the same goal.
- adaptive software
- 14%· concepts
- Differ
- 14%· companies
- Excel
- 14%· products
- Iris ten Teije
- 14%· people
- JFrog
- 14%· companies
- Other topics
- 29%
The Pipeline Is Dead - Iris ten Teije, Sky Valley Ambient Computing
WatchAI Engineer // 19:50
We turn high signal in-person events for the top AI engineers, founders, leaders, and researchers in the world into the best free learning opportunities for millions around the world here on YouTube. Your subscribes, likes, comments, speaking, attendance, or sponsorships goes a long way toward making our biz model sustainable indefinitely. We strongly believe this industry deserves a better class of community and that we know how to do this well; we just need your support.