# AI code can work while the system gets harder to change

**Summary:** Jimmy Breck-McKye warns that useful AI patches can create conflicting conventions. Reviews should ask what each change leaves for the next person.

- Canonical: https://markhuang.ai/news/ai-code-system-coherence
- Language: en
- Author: [Mark Huang](https://markhuang.ai/about)
- Published: 2026-11-16
- Section: News
- Tags: AI Coding, Software Engineering, Maintainability, Technical Debt, Decoherence
- Source: [Jimmy Breck-McKye](https://www.breck-mckye.com/blog/2026/11/software-is-breaking-down-at-a-systemic-level/)
- License: https://creativecommons.org/licenses/by-nc/4.0/

---

![Copper and blue pipes meet through a complicated cluster of carefully fitted adapters](https://cdn.markhuang.ai/news/ai-code-system-coherence/hero.webp)

*Conceptual illustration generated with OpenAI gpt-image-2: sound individual repairs can leave a system with more connections to understand.*

Replacing a leaking pipe can leave a building harder to maintain, even when the repair works. In Kristen Maevyn's [plumbing thought experiment](https://maevyn.codes/decoherence), a building gradually acquires a mixture of old and new pipe standards. The difficult question arrives later: can someone safely raise the water pressure? Answering it now means understanding the patches and the connections between them.

Jimmy Breck-McKye applies that idea to AI coding in his [November 15, 2026 essay, "Software is breaking down at a systemic level."](https://www.breck-mckye.com/blog/2026/11/software-is-breaking-down-at-a-systemic-level/) He focuses on code that does what someone asked. Each change can make sense on its own while leaving the whole system less coherent. Faster implementation could then mean accumulating integration work faster, too.

I like this as a review question: what does a patch leave behind for whoever changes the software next? Google Cloud's [2025 DORA report](https://dora.dev/research/2025/dora-report/) offers a related framing: AI amplifies an organization's existing strengths and weaknesses. That leaves room for substantial benefits when a team uses the tools to simplify its system.

## Why a good repair can complicate the building

Breck-McKye borrows the term "decoherence" from Maevyn. She defines it as the accumulation of locally valid but incompatible or redundant models within a system, until changing the system mostly involves translating between them and preventing side effects. Here, models means the different ways software represents and organizes its work. It does not mean a collection of competing AI products.

The plumbing analogy helps explain why this differs from an ordinary broken component. A sound adapter might be exactly what a repair needs today. Several generations of adapters, spread across an unfinished replacement project, make the next job harder to reason about. The building still has running water. That tells you less than you might hope about how safely it can be altered.

For software, the distinction is between whether a change works and whether it fits the surrounding design. Breck-McKye's argument is that agents can accelerate changes at the component level without preserving that larger understanding. He extends the concern to "organisational cognition," a company's ability to think and respond as a coherent unit.

The warning also applies to human work. Maevyn describes people creating decoherence under pressure. In her account, changes create it: leaving a system untouched does not, by itself, add more incompatible conventions. An agent that helps finish a migration or remove duplication could push in the other direction.

## The collapse claim needs more evidence

Breck-McKye opens with a series of outages and security incidents, then argues that software is deteriorating systemically. The essay does not establish that those incidents share an AI-related cause or provide an industry-wide measure of declining reliability. I take its account of conflicting assumptions more seriously than its forecast of a new normal.

The [Hacker News discussion](https://news.ycombinator.com/item?id=50520866) raises an objection worth keeping: an account of failures also needs to consider the software that keeps working. Other commenters question the essay's historical picture. These are reader objections, not a reliability study, but they expose a gap. A cluster of memorable failures cannot tell us the failure rate across everything software now does.

Maevyn is also explicit about the limits of her own demonstration. Her simulation uses arbitrary weights and shows how a complexity trap can emerge from simple rules. She says it does not prove that real complex systems will evolve that way. The illustration helps explain a possible mechanism; it cannot establish inevitable collapse.

Measuring the immediate speed benefit is difficult, too. In its [February 2026 productivity update](https://metr.org/blog/2026-02-24-uplift-update/), METR said selection effects and difficulties measuring time made its later developer experiment an unreliable estimate of AI's current impact. The researchers thought greater speedup was likely, but called their data weak evidence for its size. That study concerns task productivity. It does not settle what repeated changes do to maintainability.

## Review what the next change will require

My practical response is to ask a second question during review. After checking whether the patch does its job, ask which existing convention it follows and what new exceptions it introduces. If it needs a translation layer, someone should be able to explain why that boundary belongs there.

This extends the argument in [my earlier post about the cost of proving an AI patch](https://markhuang.ai/news/coding-agents-patch-proof-cost). A team pays to verify and own an accepted change. I want to examine whether that change makes later work easier to understand. A passing test is useful evidence of behavior; design review still has a job.

Different conventions can coexist during a deliberate transition. In that case, I want to know which system owns each responsibility, where the temporary connections sit, and what would let the team retire them. These are proposed review habits, not experimentally proven remedies.

The plumbing repair may still be the right repair. The question is whether the next person can explain the resulting building well enough to change its pressure safely. That is the part of Breck-McKye's warning I want to keep when asking an agent to make the next patch.
