'LLMs Are Eroding My Software Engineering Career': The Thread Every Developer Is Reading

Every few months a piece of writing escapes the niche corner of the internet where it was published and becomes the thing every developer is quietly reading at their desk. This month it was a blog post by a finance and payments engineer with a decade of domain expertise, titled around a single uncomfortable claim: LLMs are eroding his software engineering career. Within five hours of hitting Hacker News, the post had drawn 602 upvotes and 551 comments. It struck a nerve not because it was hysterical — it was the opposite — but because it was methodical. The author did not claim AI would take his job. He examined, one by one, the three professional advantages he had spent ten years building, and showed how each was being commoditized underneath him.

The Post That Stopped the Front Page

The author's credentials matter to why the post landed the way it did. He is not a bootcamp graduate worried about breaking in, and not a generalist whose work was always closest to the automation frontier. He is a specialist — ten years inside finance and payments systems, the kind of domain where regulatory nuance, settlement edge cases, and institutional scar tissue have historically functioned as a moat. If anyone was supposed to be safe from the commoditization wave, it was someone like him. That is precisely the argument he dismantles.

His structure was deliberate. He identified the three advantages that justified his compensation and seniority, and tested each against what frontier models and coding agents can now do. The first was specialized domain knowledge: the accumulated understanding of how card networks, ledgers, and reconciliation processes actually behave, knowledge that used to take years of incident response to acquire. LLMs trained on the public corpus of payments documentation, postmortems, and standards now answer a striking fraction of the questions that once required him. Not all of them — but enough to change what a team will pay for.

The second was distributed-systems debugging: the craft of tracing a failure across services, queues, and retries until the real cause surfaces. Agentic tools that can read logs, propose hypotheses, and run experiments are not yet better than a great debugger — but they have compressed the gap between a great debugger and a median one, which is the same thing as compressing the premium the great one commands. The third, and the one he found most painful, was architectural judgment: the ability to look at a proposed design and know, from experience, where it will hurt in two years. Models now produce plausible architectural reviews on demand. Whether they are right is a separate question — and that question turned out to be the heart of the comment thread.

That single sentence — the most upvoted comment in a thread of 551 — captures the new epistemic condition of the working engineer better than any think piece this year. Inside your domain, an LLM is a brilliant junior colleague whose mistakes you catch instantly. Outside it, the same model produces output of identical confidence and polish, and you have lost the ability to tell which is which. The commenter was not saying agents are bad. He was saying that the boundary of his own expertise had become the boundary of his ability to supervise them — and that everything inside the boundary was getting cheaper while everything outside it was getting riskier.

The Layoff Backdrop: Why the Despair Feels Earned

The post would not have hit 602 upvotes in five hours in a calm labor market. It landed in the middle of the worst opening quarter for tech employment in three years. Q1 2026 logged 52,050 tech layoffs — the highest first-quarter total since 2023, when the industry was still unwinding its pandemic-era overhiring. Oracle alone accounted for 30,000 of those cuts, a reduction announced with the now-standard language about reallocating investment toward AI infrastructure.

The hiring side looks worse than the firing side. Software development postings on Indeed are down 68.8% from their February 2022 peak — a collapse that has now persisted long enough that it cannot be dismissed as a post-pandemic correction. A correction reverts. This has not reverted. For an engineer reading the viral post between rounds of a job search, the author's thesis did not read as speculation. It read as a diagnosis of the market they were already living in.

It is worth being precise about what the posting data does and does not say. Postings measure advertised demand, not employment. Companies have learned to hire through networks, internal mobility, and contractor conversion in ways that never touch a job board, and the February 2022 peak was itself an anomaly — the single most overheated hiring month in the history of the industry. Measuring from the top of a bubble exaggerates the fall. But even with those caveats, a 68.8% decline over four years describes a market in which the default experience of looking for work has fundamentally changed.

The Contradiction: Employment Is at a Record High

Here is where the story stops being simple. While Q1 2026 was producing the worst layoff numbers in three years, global developer employment was setting a record. Developer employment grew 3.8% in 2025 — roughly 72,000 net new jobs — bringing the global total to 28.7 million, the highest figure ever recorded. Both facts are true at the same time, and any honest account of the moment has to hold both.

How do record employment and record despair coexist? Three mechanisms, none of them comforting in the way the headline number suggests. First, the growth is not where the layoffs are. Net job creation is concentrated outside the large US tech companies whose cuts dominate the news — in non-tech industries hiring software talent, in markets outside the United States, and in smaller firms that never overhired. The engineer laid off from Oracle and the engineer hired by a logistics company in another country are both in the data; they are not in the same life.

Second, the composition of the work is changing faster than the headcount. Job postings requiring AI tool experience are up 340% year over year, while postings for pure implementation roles — the take-a-ticket, write-the-code, close-the-ticket jobs — are down 17%. The market is not deleting software engineers. It is re-pricing a specific kind of software engineering, the kind whose output most resembles what an agent produces. That is exactly the work the viral post's author identified as the eroding portion of his own value.

Third, the entry point has narrowed dramatically even where the aggregate has grown. The net growth in developer employment is overwhelmingly a story about experienced engineers being retained and redeployed; the rungs at the bottom of the ladder are the ones being removed. We documented this in detail in our analysis of how junior developer jobs fell 20% while their seniors got raises. A market can set an employment record and still be a terrible place to start a career — or, as the viral post argues, a destabilizing place to be mid-career in a commoditizing specialty.

Reading the Thread: What 551 Comments Actually Argued About

The comment thread split along a fault line that is worth naming, because it is the same fault line running through every engineering organization right now. One camp read the post as overdue realism: the profession spent two decades treating accumulated knowledge as an annuity, and the annuity is being repriced. The other camp pointed at the employment data and the long history of premature obituaries for programming — COBOL automation, CASE tools, offshoring, no-code — and argued that every wave of commoditization has so far expanded the total demand for software and the people who can be accountable for it.

Both camps were arguing past the post's actual claim. The author never said software engineering was ending. He said his particular portfolio of advantages was depreciating, and he is probably right. The honest synthesis of the thread is that commoditization is real, uneven, and survivable — but only for engineers who can identify which of their skills are depreciating and reinvest deliberately. The skill of "knowing things an LLM also knows" is depreciating at the speed of each model release. The skill of "being accountable for what ships" is not.

There is also a behavioral dimension the thread mostly skipped: the same engineers worried about erosion are accelerating it through their own tool adoption. As we covered in our analysis of the METR findings on developers who now refuse to code without AI, dependence on agents has become near-universal among working engineers even where measured productivity gains are ambiguous. Every engineer who routes their domain knowledge through an agent is also, in a small way, training the organization to believe the agent was the source of it.

What Is Actually Being Commoditized — and What Is Not

The most useful way to read the viral post is as a sorting exercise. The author's three advantages are not equally exposed, and the differences between them generalize to almost every engineering specialty. Knowledge that exists in public text — documentation, standards, postmortems, Stack Overflow — is the most exposed, because it is literally the training data. Judgment that requires organizational context the model has never seen is the least exposed. Most engineering careers are a blend, and the blend is what determines the slope of the erosion.

Notice what the right-hand column has in common: every item is a form of verification or accountability. The top comment's lament — "I can no longer call BS on the agents" outside my domain — is also, read in reverse, a market signal. The ability to call BS is becoming the scarcest skill in software, because agents have made the production of confident output free while leaving the cost of wrong output exactly where it was. Someone still pays for the wrong output. The engineer who can prevent that payment is the one whose value compounds.

This is also why the 340% growth in AI-skills postings should not be read as "learn to prompt." Prompting is the trivially copyable part. What employers are actually trying to buy — often without being able to articulate it — is engineers who can run agentic workflows and stand behind the result: who know which outputs to trust, which to test, and which to throw away. The interview process has been slower to adapt than the job itself, a gap we explored in our analysis of how AI broke the coding interview; companies are still screening for the depreciating column while paying premiums for the compounding one.

The Sentiment Slide Is Rational

One more data point deserves attention because it reframes the despair as something other than fear: developer favorability toward AI tools fell from 77% in 2023 to 60% in 2026. The naive reading is backlash. The better reading is calibration. In 2023, most developers had used the tools for months; by 2026 they have used them for years, through enough failure modes to price them accurately. Favorability falling while usage approaches universality is not a contradiction — it is what a workforce looks like when it has stopped evaluating a tool as a novelty and started evaluating it as a colleague, employer, and competitor all at once.

What to Actually Do: A Mid-Career Playbook

If the viral post describes your situation — deep in a domain, watching the moat drain — the worst response is the most common one: doubling down on accumulating more of the same knowledge. The depreciation curve does not care how much of a depreciating asset you hold. The useful moves all involve shifting weight from the left column to the right.

First, move toward the failure surface. Systems still fail in ways that are not in any training corpus, and the engineers who own incident response, reliability, and the consequences of failure are accumulating exactly the context models cannot have. Second, make your verification ability legible. "Reviewed and approved 400 agent-generated changes with a 2% defect rate" is a résumé line that did not exist three years ago and will be among the most valuable ones three years from now. Third, attach yourself to outcomes rather than components. Domain expertise in payments depreciates; accountability for a payments system that moves real money does not, because accountability is the one thing organizations cannot delegate to a model — as every attempt to put an agent's name on an incident report has so far demonstrated.

And for the engineers a decade behind the post's author: the entry-level math is harsher, but the prescription is the same in kind. The career ladder's missing rungs are real, the pipeline damage is real, and the compounding skills are the same verification-and-accountability skills — they are just harder to acquire without the production scar tissue that used to come with junior roles. That is the genuinely unsolved problem in this story, and it belongs to employers as much as to graduates.

Conclusion: The Erosion Is Real. So Is the Riverbank.

The viral post deserved its 602 upvotes. It is the most honest first-person account yet written of what commoditization feels like from inside a career that was built correctly by every rule that held until 2023. The despair in the thread is not irrational, and waving the 28.7 million employment record at it misses the point — record employment in a profession whose internal economics are being violently re-sorted is cold comfort to the person on the wrong side of the sort.

But the data also refuses the simple doom reading. A market adding 72,000 net jobs while paying a 340% premium for a new skill profile is not a dying market. It is a market mid-rotation — punishing accumulated recall, rewarding verification and accountability, and doing both faster than any prior transition in the industry's history. The author of the post is right that his three advantages are eroding. The engineers who fare best over the next five years will be the ones who read that essay not as an obituary but as an asset audit, and rebalance accordingly.