•Jesse Landry
The Velocity Trap
Something structural is happening in software, and most commentary is still aimed at the wrong layer. The dominant narrative says AI is eliminating the need for engineers. Models generate code....
FrameworksTop News
Something structural is happening in software, and most commentary is still aimed at the wrong layer. The dominant narrative says AI is eliminating the need for engineers. Models generate code....
Something structural is happening in software, and most commentary is still aimed at the wrong layer.
The dominant narrative says AI is eliminating the need for engineers. Models generate code. Features appear in hours. Entire applications can be scaffolded in an afternoon. The conclusion feels efficient and modern: if code generation accelerates, the talent shortage dissolves.
That conclusion mistakes production for constraint.
Software has always operated on multiple clocks. The production clock governs how fast new code can be created. The verification clock governs how fast that code can be reviewed, tested, and understood. The accountability clock governs how fast failures are discovered, audited, and assigned responsibility. AI has made the first clock exponential. The other two remain stubbornly slower.
When one clock accelerates inside a system governed by slower clocks, pressure accumulates. In engineering, that pressure has a name.
Technical debt.
For decades, what Ward Cunningham called technical debt was largely intentional. Teams knowingly cut corners to ship sooner, accepting that they would repay the cost later through refactoring. That model assumed humans were the ones deciding to borrow.
AI introduces a more dangerous variant. Not shortcut debt. Overflow debt. Debt created when the rate of system change exceeds the organization’s capacity to form accurate beliefs about what the system now depends on.
In a controlled trial conducted in 2025, experienced open source developers working on large, familiar codebases completed tasks 19 percent slower when AI tools were allowed, while believing they were 20 percent faster. The gap between perceived velocity and actual velocity is not a rounding error. It is measurable misalignment. Production accelerates in appearance. Verification load increases in reality.
Comprehension thins.
Everything compiles. Tests pass. The feature deploys. Documentation improves. Yet the surface area of the system expands faster than anyone can mentally model it. Dependencies multiply. Interfaces grow more brittle. Edge cases accumulate at the seams.
Security exposes this imbalance most clearly. In 2025, Veracode tested more than one hundred large language models across eighty coding tasks and found that 45 percent of generated outputs introduced OWASP Top Ten vulnerabilities. Larger models did not materially reduce the rate. The production clock improved. The accountability clock did not accelerate at the same pace.
The result is not immediate failure. It is delayed fragility.
Global analysis of ten billion lines of code across forty seven thousand applications found sixty one billion workdays of accumulated remediation effort already embedded in existing systems. Nearly half of the world’s code was classified as structurally fragile before AI multiplied production capacity. Overflow debt does not replace that baseline. It compounds it.
The system absorbs this pressure through verification.
When production outpaces comprehension, verification becomes the bottleneck. Code must be reviewed. Architectures must be validated. Security boundaries must be audited. Compliance obligations must be mapped. The verification clock becomes the binding constraint.
And when verification becomes the constraint, the scarce resource is no longer code generation capacity. It is stewardship.
This is where the talent narrative reenters, not as a pivot but as a consequence.
Overflow debt increases verification load. Verification load concentrates on senior engineers, architects, and system owners. Those individuals are finite. They cannot scale at the rate of production. Meanwhile, the apprenticeship layer that historically produced them is thinning. New graduate hiring at large technology firms has fallen by roughly fifty percent from pre pandemic levels, and entry level roles most exposed to automation continue to contract. The pipeline that produces future stewards narrows precisely when stewardship becomes existential.
The shortage has not disappeared. It has moved.
There will be no scarcity of prompts. There will be scarcity of judgment.
DORA’s multi year research confirms the pattern. As AI adoption increases, local metrics such as documentation quality and code quality show measurable improvement. Throughput may rise. Yet delivery stability declines. The production clock accelerates. Coherence does not automatically follow.
The ecosystem is mistaking output for capability.
An application built in a weekend is not equivalent to a system that can survive a decade of scale, audit, migration, and regulatory scrutiny. Enterprises do not fail because they shipped too slowly in sprint one. They fail because invisible coupling surfaces under load, because compliance timelines collide with undocumented logic, because capital expects liquidity from systems that cannot evolve safely.
AI does not remove entropy from software systems. It increases the rate at which entropy is introduced.
This is not an argument against AI. It is an argument about physics.
When exponential production is injected into institutions governed by slower verification and accountability clocks, fragility compounds invisibly before it becomes visible. Technical debt is the engineering expression of that mismatch. Talent scarcity is the human expression of the same force.
The constraint in software was never typing speed.
It was always the rate at which organizations could understand what they had built.
AI is not eliminating the need for talent.
It is eliminating the illusion that production was the hard part.
Stewardship remains the bottleneck.
And stewardship does not scale at the speed of generation.