The systems we built to remove friction are now removing the structures that keep them honest.

The Pattern

The week opened with a question about practice. March 16's analysis examined the Fowler study: 158 engineers using AI produced 20% more code, but their approval rates dropped. The distinction that emerged was between construction offloading and retrieval offloading. One preserves the practitioner. The other hollows them out. That finding set the tone for everything that followed.

By midweek, the pattern scaled from individual engineers to entire organizations. March 18's analysis showed the DORA report confirming that 90% of developers now use AI tools without automatic improvement in delivery performance. The same day, an AWS outage exposed the gap between stated architecture and actual architecture. March 19's analysis sharpened the point further. A Cloudflare engineer rebuilt a Next.js equivalent in a week. Generation has become a commodity. The scarce resource is now judgment, the ability to evaluate what was generated and decide whether it belongs. Teams without evaluation infrastructure are producing more and understanding less.

The second half of the week revealed where this trajectory leads when no one intervenes. March 20's analysis broke down the Cursor Composer 2 incident, where developers discovered their tool was secretly routing through Kimi 2.5, a Chinese language model. The tool optimized for speed and undermined its own trust layer in the process. March 21's analysis widened the lens to institutions. The Pentagon designated Anthropic a "supply-chain risk," not because the technology failed, but because verification cannot keep pace with capability. And March 22's analysis landed the week's thesis directly. A peer-reviewed study titled "Against Frictionless AI" argued that the friction we keep trying to eliminate is structurally necessary. Friction is not waste. It is load-bearing infrastructure. Remove it without understanding what it held, and the system collapses under its own output.

The Tension

### AI & Agents

The dominant signal this week is that AI capability is outrunning the verification systems designed to govern it. The Fowler study, the Cursor incident, and the Pentagon's Anthropic designation all point to the same structural problem. Capability scales automatically. Trust does not. The 22,511 agent skills audit from Saturday suggests the surface area for unverified behavior is growing faster than anyone is tracking.

### Dev Infrastructure

DORA's findings confirm what the best teams already know: tooling adoption is not the bottleneck. Workflow design is. The Cloudflare rebuild story illustrates that generation cost is collapsing toward zero, which means the competitive layer has shifted entirely to evaluation, integration, and decision quality. Teams still measuring productivity by output volume are optimizing the wrong metric.

### Geopolitics & Power

The Cursor-Kimi 2.5 supply chain breach and the Pentagon's Anthropic designation are two faces of the same concern. Geopolitical boundaries are now embedded in software dependencies. FedRAMP authorization conversations have accelerated. The question is no longer whether AI tooling has geopolitical implications. It is whether procurement processes can adapt before the next incident forces their hand.

### Human Performance

The Fowler study is the most important human performance signal of the quarter. Construction offloading degrades the neural pathways that produce expert judgment. Retrieval offloading preserves them. This is not a productivity question. It is a formation question. The practitioners who understand this distinction will compound skill while their peers plateau. The "Against Frictionless AI" study reinforces this at the systemic level. Energy buffers, cognitive friction, deliberate slowness. These are not inefficiencies to optimize away. They are the substrate of durable performance.

### Energy & Regulation

NRC oversight subordination surfaced Saturday as a quiet but significant signal. Regulatory friction in the energy sector is being reduced in the name of speed. The pattern mirrors what is happening in software. When the constraint layer is treated as the enemy, the system loses its ability to self-correct.

What This Unlocks

Every pillar this week told the same story from a different angle. The friction layer is being dismantled, not because it failed, but because it is slow. And slowness has become intolerable in systems optimized for throughput. But friction was never just delay. It was where verification happened. Where judgment formed. Where trust was built.

I think the next six months will sort organizations into two categories. Those that understood friction was structural and preserved it deliberately. And those that optimized it away and cannot explain why their systems are failing despite record output. The second group will be larger. The first group will be more durable.

Watching Next

**CONFIRMED:** AI supply chain opacity is now a national security conversation. The Cursor-Kimi 2.5 incident and Pentagon-Anthropic designation both validate what Week 10 flagged about model provenance becoming a procurement requirement.

**STILL WATCHING:** Whether DORA's next quarterly data shows any correlation between AI tool adoption and incident rates. The current data says no improvement in delivery. The question is whether it shows active degradation.

**STILL WATCHING:** FedRAMP authorization timelines for major AI vendors. The gap between commercial adoption speed and federal certification speed is widening.

Underweighting

Week 11 predicted that developer tooling trust would become a first-order concern before Q2. The Cursor incident confirmed this two weeks ahead of schedule. The mechanism was different than expected. Not a data breach, but a silent model substitution. The implication is the same. Developers are now running forensics on their own tools.

Bottom Line

Week 13 should watch for organizational responses to the verification gap. The pattern is clear. The question is who moves first to formalize evaluation infrastructure, decision records, and friction-preserving workflows. Watch for the first major enterprise to publicly reintroduce deliberate slowness as a competitive strategy. That will be the signal that the thesis has crossed from analysis into practice.

Sources

1680 articles scanned / 71 sources