Will AI replace developers?
Tung Bui Thanh (MikeDaDev)11 mins read

On this page
- # what the "replacement" argument actually claims
- # what Hoang told his engineer
- # "AI writes 90% of the code" is not "AI does 90% of the work"
- # productivity: real gains, but nobody can measure how big
- # the market: where the replacement side is partly right
- # where I push back, on both sides
- # face it and adapt
Lately the same sentence comes up everywhere: "AI is about to replace developers." Colleagues say it half-jokingly over lunch, friends outside tech ask about it after reading the news, and every week there is a new headline along the lines of "AI now writes most of the code".
Hoang Nguyen (codeaholicguy) just wrote about a 1:1 in which an engineer on his team asked exactly that, because Claude Code now does most of the work they used to do by hand. His essay is worth reading and I summarise it below. But this question deserves an answer from data rather than from feelings, so this post does three things: break the "replacement" argument into separate claims, set each claim next to primary data, and point out where the replacement side is right, because some of it is.
# what the "replacement" argument actually claims
Collect the statements and the "AI will replace developers" side usually rests on four ideas:
- AI will soon write almost all code. In March 2025, Anthropic CEO Dario Amodei said that within three to six months AI would write about 90% of code, and within about 12 months essentially all of it.
- AI can do an engineer's job. In January 2025, Mark Zuckerberg said on Joe Rogan's podcast that Meta and other companies would have AI working as a mid-level engineer that same year.
- The market is already shrinking, especially for people entering the profession.
- Developers feel it themselves: "the agent already does most of my work."
The fourth point worries people most, because it comes from experience rather than headlines. So I start with Hoang's essay.
# what Hoang told his engineer
What I told an engineer who asked if AI will replace us was published on Substack on 16 November 2026 (originally on codeaholicguy.com on 6 November). His short answer: the way we execute is changing a lot, but the core of the job is not. The main points:
- The job is problem-solving. Someone has a problem, the team works out what the real problem is, decides what is worth building, and writes software that takes load off someone's life. Consumer apps, enterprise software and internal tools all work this way.
- Fred Brooks drew this line in 1986. In "No Silver Bullet", Brooks split difficulty into essence (deciding precisely what to build) and accident (tooling friction). High-level languages and better IDEs attacked the accidental side. Hoang argues agents are the biggest attack on that same side so far, and they do not touch the essence.
- The bottleneck moves. Apply Amdahl's law to a team: speed up one phase and the rest sets the pace. Work piles up in review and verification. "I automated most of my work" and "I deliver more value" are different claims.
- History follows a pattern. Lamplighters, ice cutters, switchboard operators, elevator operators, human computers: new tools hit hardest the people tied to the old tool, but the need stays. For developers, typing code by hand is the old tool; the need is the problem being solved.
- An 80/20 thought experiment. Assume an agent can do 80% of your work and put your effort into the other 20%: understanding an unclear problem, choosing an approach when every option has a cost, owning whether it works (define done as the outcome for the user, not a merged pull request), and caring whether the product fits into people's lives.
- What he cannot promise. The entry path is narrowing, judgment is harder to build when the usual path shrinks, and AI might one day do the 20% as well.
I agree with most of it. The best part is that he does not offer empty reassurance: he admits it is "honestly depressing" to watch a machine do what took him years to learn, and he is upfront about what he does not know. Now let us check each claim against data.
# "AI writes 90% of the code" is not "AI does 90% of the work"
Amodei's forecast has been scored many times. Some call it false if you measure code actually merged at large companies; others call it true if you only count lines generated. Arguing about that number misses the point, because lines of code were never the measure of an engineer's work.
The best data on who decides what comes from Anthropic itself. In June 2026 it analysed about 400,000 Claude Code sessions from about 235,000 people (October 2025 to April 2026). People made about 70% of planning decisions but only about 20% of execution decisions. In other words, the agent does most of the typing, while deciding what to build and how still sits with people.
The report has a more striking detail: verified success was about 15% for novices, against 28 to 33% for intermediate and above. Novices also abandoned troubled sessions far more often (19% against 5 to 7%). Anthropic notes these results are preliminary, since a model classifies the transcripts. But the direction is clear: expertise does not lose value with agents, it decides whether the agent produces a result at all.
# productivity: real gains, but nobody can measure how big
If AI could replace developers, it should at least make developers measurably faster. The evidence here is messy.
The most cited study is METR's 2025 randomised trial: 16 experienced open-source developers worked 246 real tasks in their own repositories. With AI allowed, tasks took about 19% longer, while the developers themselves believed AI had made them about 20% faster. That gap between perception and measurement is the study's biggest lesson.
In February 2026 METR published a second round with 57 developers and more than 800 tasks. The estimates lean towards a speedup, about 18% for returning developers and 4% for new ones, but both confidence intervals cross zero. The reason METR calls the new data unreliable is telling: 30 to 50% of developers said they chose not to submit some tasks because they did not want to do them without AI. That is indirect evidence that AI really helps, and also evidence that nobody has measured how much.
The DORA 2025 report (nearly 5,000 practitioners) looks at teams: higher AI adoption is now associated with higher delivery throughput, but also with higher delivery instability. DORA calls AI an amplifier: good teams get better, fragmented teams expose their weaknesses faster. The Stack Overflow 2025 survey found 84% of developers use or plan to use AI, while 46% do not trust the accuracy of its output, up from 31% a year earlier. The top reason developers would still ask a person is not trusting AI's answers.
Put together: the tools are useful, everyone uses them, and people still own the result. That describes a tool, not a replacement.
# the market: where the replacement side is partly right
This is the part both sides most often quote wrongly, so it is worth being precise.
| Source | Figure | Important caveat |
|---|---|---|
| Stanford Digital Economy Lab (Aug 2026) | Employment of 22 to 25 year olds in highly AI-exposed jobs is about 19% below where it would otherwise be (15% in 2025) | A gap against less-exposed peers, not a raw decline. Experienced workers show no comparable gap. The authors call it descriptive, not causal |
| BLS (Aug 2026) | Software developer jobs projected to grow 10% from 2025 to 2035; "computer programmers" to decline 7% | Growth is lower than the previous projection. BLS rates the occupation "very high" on AI exposure |
| Indeed Hiring Lab (Nov 2026) | Software development postings still about 19% below February 2020, but up 23.8% year over year | The low point dates from the correction after the 2021 to 2022 hiring boom, before agents were common |
Read carefully, the picture is not "AI is wiping out programming". Stanford says plainly that it does not see widespread, economy-wide displacement, and that the main mechanism is companies hiring fewer young people, not more layoffs. Postings are recovering and BLS still projects growth.
It is also not true that nothing is happening. Three points where the replacement side is right:
- "Translate a spec into code" work really is shrinking. BLS projects computer programmers, the people whose job is mainly writing code to a specification, to decline 7%. That part of the work is being replaced.
- Newcomers are hit hardest. The 19% gap for 22 to 25 year olds grew from 2025 to 2026. For a new graduate, "the profession will survive" is cold comfort.
- Some companies will need fewer engineers for the same amount of work. Hoang acknowledges this too.
# where I push back, on both sides
The lamplighter example is a good one, but it has a side people skip: the need for light stayed, and the lamplighters still lost their jobs. "The profession survives" and "you keep your job" are different sentences. So I think the right question is not "will AI replace developers" but "which part of what I do will be replaced, and how fast am I moving".
The replacement side, in turn, tends to forget an old economic rule: the Jevons paradox. When something gets much cheaper, people usually use more of it, not less. When software gets cheaper to build, more problems become worth writing software for: internal tools nobody built because there were never enough people, processes that still run on spreadsheets and email. This year's 23.8% rise in postings does not prove that, but it runs directly against the "work is about to run out" story.
There is also a practical lesson from outside software. In 2024 Klarna boasted that its AI assistant did the work of 700 customer service agents. By May 2025 its own CEO said the automation push had "gone too far" and hurt quality, and the company started hiring people again under a flexible model. Klarna still calls itself AI-first, but the story shows that replacing people with AI is much easier than keeping quality after you have. In software, where one bug can take down a payment system, the cost of poor quality is higher still.
# face it and adapt
First, face it. The typing really is being replaced, the way into the profession really is narrowing, and the sting of watching a machine do what took you years to learn is real too. Denial ("AI is just fancy autocomplete") and panic ("start planning a career change") are both easier than looking straight at it, and neither helps you decide what to do next. Look straight at it and you see that the typing is being replaced, while understanding the problem, choosing an approach and owning the result are not, and the current data supports that.
Then adapt, with concrete actions rather than worry:
- Get really good with agents, as Hoang advises. Avoiding the tools does not protect anyone's job. But watch where the saved time goes: into understanding the problem, choosing an approach and verifying, or into producing more of the same faster.
- Take ownership of verification. Write tests, set up an environment where agents can run and check their own work, and read risky changes carefully. Every report points to this as the slowest and most valuable part right now.
- If you are new: build judgment on purpose. Do not let the agent solve, from the start, problems you have never solved yourself. Do it once on your own, then compare with how the agent does it and ask why they differ. The data shows novices succeed with agents half as often, so that gap has to be closed deliberately.
- Measure outcomes, not pull requests. For teams and individuals alike. 300 PRs a week that users do not notice still add up to zero.
As for "what if AI does the remaining 20% one day", I cannot answer that, and I doubt anyone can with confidence. But if that day comes, as Hoang writes, the questions will be much bigger than programming. Until then, the most valuable work is still the work that needs a person to own it.
For more on working with agents without wrecking the codebase, see the summary of Lauren Tan's (SpaceX) All Code Is Slop and the post on ADLC.
References:
- Hoang Nguyen (codeaholicguy): What I told an engineer who asked if AI will replace us
- Anthropic: Agentic coding and persistent returns to expertise (June 2026)
- Stanford Digital Economy Lab: AI employment gap for young workers widened to 19% (August 2026)
- METR: developer productivity study (2025) and experiment design update (February 2026)
- DORA 2025: State of AI-assisted Software Development
Just personal notes, shared in the hope they are useful. If you found this helpful, feel free to pass it on to your friends and colleagues!
Happy coding ๐ ๐๐ป ๐ ๐ฅ.