Skip to content
Guides

October 24, 2026

Is Manual Coding Dead? What History Says About the End of Hand-Written Software

14 minute read

Manual coding is disappearing: agents already write 47% of professional code. What the vanished job of the human "computer" — and the people who became the first programmers — say about where the work goes next.

Is Manual Coding Dead? — BrowserStay blog card

Is manual coding dead? It is the question every developer is asking in 2026, and the honest answer starts with a job title most people have forgotten: for most of the twentieth century, "computer" was a person — usually a woman with a mathematics degree — who calculated trajectories and statistics by hand. That job died when machines took over the arithmetic. The work did not die with it: the people became the first programmers, and computation exploded into everything.

This research brief follows that precedent through today's data. Roughly 47% of professional code is now written entirely by AI agents, yet AI pull requests are far less likely to merge than human ones, and economists still argue about whether the ATM or the spreadsheet is the right map for what happens next. Every figure below comes from a primary source — NASA's own histories, NBER working papers, JetBrains' 15,000-developer survey — all linked in the Sources list at the end. (For the longer arc behind today's agents, see The Story of AI: From Turing's Question to Autonomous Agents.)

The verdict up front

  • The share of professional code typed by hand is collapsing: developers now report that roughly 47% of their code is written entirely by AI agents, 38% is written with AI assistance, and only about 27% is written fully by hand.
  • History offers a precise precedent: "computer" was a job title for humans who calculated by hand. That job died in the 1970s — and was replaced by something larger, not nothing.
  • The economic pattern that governed the ATM and the spreadsheet — automation of tasks creating more work than it destroys — has held for two centuries, but it is not a law of nature, and the last four decades show displacement getting stronger while job-creating augmentation gets weaker.
  • The honest answer to "is manual coding dead?": hand-writing every line as a job's core activity is ending fast; coding — deciding, specifying, verifying, owning software — is not dying, it is moving up a level, exactly the way calculation moved from the computer pool to the programmer.

1. When the computer wore a skirt

Before there were electronic computers, a computer was a person. At Langley Memorial Aeronautical Laboratory, the first "computer pool" opened in September 1935 with five women, most with degrees in mathematics or science, who read wind-tunnel data and calculated trajectories by hand. The profession itself was older — human computers had worked in astronomy, statistics, and ballistics since the late nineteenth century.

By 1946, Langley's computing operation had trained about 400 women. At the Jet Propulsion Laboratory, a team of women known as the human "computers" calculated the positions of spacecraft from telemetry — work that began in 1936 and continued into the space race. As the Smithsonian puts it: just before the digital age emerged, computers were humans, sitting at tables doing math laboriously by hand, powering everything from astronomy to war and the race into space.

Then the machines arrived, and the job title died. But watch what actually happened at the boundary. The human computers did not simply disappear; at NASA and JPL many of them took on programming duties — they became the first computer programmers, the people who told the new machines what to calculate. The occupation of hand calculation died; the work of computation exploded, and the humans moved one level up the stack. Katherine Johnson's team checking a Mercury trajectory and an engineer writing Fortran for the same mission were doing, in substance, the same job at different altitudes.

That is the template this report applies to coding.

2. The numbers in 2026: manual coding is disappearing, fast

The best current measurement comes from JetBrains' Developer Ecosystem Survey 2026, which asked over 15,000 professional developers worldwide what share of their work code was written how, between May and July 2026.

  • About 47% of code is now fully written by AI agents; about 38% is written with AI assistance; about 27% is written fully by hand.
  • More than half of all developers now write less than 20% of their code manually, and one in five writes no code without AI help at all. Fully agentic developers — over 80% agent-generated code — are still a minority, at about 22%.
  • The JetBrains researchers split developers into three profiles: agentic coders (~31%, averaging 84% agent-generated code), AI-assisted coders (~47%), and manual coders (~23%, still writing about 75% by hand).
  • Senior developers lead the shift: about a quarter of seniors now generate over 80% of their code with agents, more than juniors do.

The surrounding data agrees that the direction is one-way. Ninety percent of professional developers used AI coding agents at work at least weekly in mid-2026, and 68% used them daily; Claude Code's work usage jumped from 18% in January 2026 to 39% by mid-year. Google disclosed in April 2026 that 75% of its new code is AI-generated and then approved by engineers — a figure that has tripled across three disclosures since late 2024. On SWE-bench Verified, the standard benchmark for autonomous software engineering, the best systems went from 1.96% to 78.4% between October 2023 and April 2026.

But the same data shows why "dead" is the wrong word for the profession. AI pull requests at scale merge within 30 days only 32.7% of the time, against 84.4% for unassisted work; agentic pull requests wait 17.6 hours at the 75th percentile just to be picked up for review, versus 3.4 hours for human-written ones. The constraint has moved from writing code to reviewing and trusting it — which is precisely where the human work is relocating.

Who writes the code in 2026: 47% written by AI agents, 38% written with AI assistance, 27% written fully by hand
Who writes the code, 2026 — JetBrains Developer Ecosystem Survey 2026, 15,000+ professional developers, May–July 2026. Averages are calculated from answer-bucket midpoints.

And the productivity story is genuinely unsettled. METR's randomized early-2025 trial found experienced open-source developers actually took 19% longer with AI tools; its late-2025 follow-up pointed toward speedups but METR itself calls that data unreliable, partly because developers withheld tasks they refused to do without AI. Veracode's July 2026 report put the average security pass rate of AI-generated code at 56% across 100+ models, and a January 2026 Sonar survey found only 48% of developers always verify AI code before committing. Cheap generation with expensive verification is not a dead profession; it is a profession whose center of gravity moved.

3. The precedent: automation kills tasks, and mostly creates work

The strongest-provenance evidence on what happens next comes from labor economics rather than tech commentary.

MIT's David Autor and coauthors built an eight-decade database of new job titles linked to US Census microdata. Roughly 60% of employment in 2018 is in job titles that did not exist in 1940 — among professionals, 74%. The economy did not primarily shuffle people into old roles; it invented new ones.

Daron Acemoglu and Pascual Restrepo's framework names the two forces in play. Automation has a displacement effect: capital takes over tasks people used to do, always cutting labor's share. It is counterbalanced by a reinstatement effect: the creation of new tasks in which labor has a comparative advantage, which always raises labor demand. Every wave of automation is a fight between those two terms.

The case studies of the second force are well documented:

  • The ATM and the teller. ATMs cut the tellers needed per urban branch from 20 to 13 between 1988 and 2004 — but made branches so much cheaper to run that urban branch counts rose 43%, and teller jobs did not fall; the job shifted from cash handling to relationship banking. In 1970 there were about 250,000 bank tellers in the US; the number roughly doubled by the late 2000s even as ATMs passed 400,000.
  • The spreadsheet and the accountant. Around 299,000 Americans worked as accountants, bookkeepers, or auditors in 1979, as personal finance software arrived; by 1989 the field had grown 75% to about 524,000, and it stands around 1.28 million today. Cheaper arithmetic did not shrink the profession — it made financial analysis so cheap that demand for it multiplied.
  • The tractor and the farm. Self-employed and family farmworkers fell from 7.6 million in 1950 to 2.06 million in 2000, a 73% reduction, while the share of the US workforce in farming fell from 41% in 1900 to 2% by 2000 — and the economy absorbed the workers into sectors the tractor's savings made possible.
  • The power loom and the weaver. Power looms automated 98% of the labor needed to weave a yard of cloth in the nineteenth century; cheaper cloth raised demand so much that weaving jobs increased, and weavers' remaining skills grew more valuable.

Acemoglu and Restrepo's own data quantifies the engine: about half of US employment growth between 1980 and 2015 came from occupations whose titles or tasks changed — new work, not old work redistributed.

4. When the pattern breaks

This history is genuinely encouraging and genuinely incomplete, and honest research has to say both.

First, the counterexamples are real. Bessen's survey of occupational data found only one major group being persistently eliminated in developed economies — manufacturing — with its losses offset by growth elsewhere. Acemoglu and Restrepo's decomposition shows the displacement effect accelerating over the last three decades while the reinstatement effect weakened — slower employment growth is accounted for by exactly that combination. Their conclusion is not "everything works out"; it is that the historical balance was contingent, and it has been tilting.

Second, Acemoglu and Johnson warn of "so-so automation": technologies that displace workers without generating enough new productivity or new tasks to replace the demand they destroy — the case where the historical pattern genuinely fails.

Third, the teller story has a second act that matters for coding. The ATM did not kill the teller; the smartphone did. Bank branches peaked at roughly 99,550 in 2009, then fell below 78,000 by 2023 as mobile banking removed the reason customers visited at all; the Bureau of Labor Statistics now projects teller employment, about 347,400 in 2024, to decline 13% through 2034. The ATM automated what tellers did; the iPhone automated why they existed. Task automation expands adjacent roles; purpose automation eliminates them. By that framing, software development today is still in the ATM phase — agents write code, tests, and documentation, while humans own requirements, judgment, and deployment — and the open question is whether and when a second technology arrives that removes the reason the role exists.

The evidence that AI can create new work, not just displace it, is not automatic either: Autor's team finds that augmentation innovations causally boost occupational labor demand while automation innovations depress it, and that over the last four decades the demand-eroding effect has intensified while the demand-raising effect has not. Whether coding follows the spreadsheet or the teller's second act depends on which side of that ledger agentic coding lands on.

5. Where the brain goes: the abstraction ladder

Here is the optimistic case, and it is structural rather than wishful. Every major abstraction in computing's history has followed the same arc: the manual layer dies as a job, and the industry above it grows.

Hand calculation gave way to machine code. Machine code gave way to assemblers, assemblers to compilers, compilers to high-level languages, and each step was mourned as the death of "real programming" — and each step multiplied the number of people building software and the ambition of what they built. Karpathy's framing names the current step: Software 1.0 is code written by people, Software 2.0 is neural networks trained on data, and Software 3.0 is programming by natural-language prompt — "the hottest new programming language is English."

The physics of the historical pattern is cost and demand. When the computer pool automated an engineer's morning of arithmetic, engineering did not shrink; the number of research questions worth asking exploded, and the humans moved to the questions. When the spreadsheet made analysis free, clients did not want fewer analyses. When code generation approaches free, the binding constraint becomes the supply of ideas, judgment, and trust — and each of those is human work. This is the argument in its sharpest form: releasing brains from syntax is how a civilization goes to Mars, because the Mars mission is precisely what happened when humans stopped doing the arithmetic and started doing only the deciding.

But abstraction is not automatic salvation, and the current data shows the friction points. The research frontier itself has moved there: a 2026 position paper argues the bottleneck in coding agents is no longer task-solving capability but how users communicate with, supervise, and trust the systems — alignment, verifiability, steerability. A survey of the agentic-software-engineering literature names "the economics of attention" and "skill redistribution" among the five open problems that will decide whether the transition is net-positive for the discipline, and reports that 49% of jobs sampled by Anthropic in 2026 already saw AI used for at least a quarter of their tasks. The review-queue numbers in section 2 are the same phenomenon measured in pull requests.

The new human work is already visible in job shape even before it shows up in statistics: specification, orchestration, verification, and security review of machine-written systems. Senior developers — the people whose judgment is most expensive — are handing over the most code-writing, which is evidence that what they keep is the part only they can do.

6. So, is manual coding dead?

Say it precisely and the answer holds up:

  • Dead: manual coding as an activity you are paid for by the hour. Nearly half of professional code is already agent-written, more than half of developers write under 20% by hand, and every trend line points the same way. Within this decade, "I write code by hand" as a job description will read like "I calculate trajectories by hand" — a phrase from a museum.
  • Not dead: coding as directing, verifying, and owning software. The demand for software has never once been limited by the supply of people willing to type it; it is limited by cost and trust. Agents attack the cost; humans hold the trust. The profession's center of gravity moves up the stack, as it moved from the computer pool to the programmer, from the machine room to the IDE.
  • Genuinely unknown: whether the net jobs follow. Two centuries of evidence say task automation creates more work than it destroys, and 60% of today's job titles did not exist in 1940. But the same research shows displacement accelerating against weakening reinstatement, and history's averages were cold comfort for the typist whose specific job evaporated. The aggregate pattern is a tendency to be engineered toward — via training, task design, and where firms point the agents — not a guarantee.

The human computers are the right image to keep. Their job title is gone; their work is now the substrate of civilization. When the typing leaves coding, coding does not die. It does what calculation did in 1970: it stops being the job and becomes the medium.

Sources

Tags:#AI#Career#Future#Technology
Share this post:X / TwitterLinkedInRedditWhatsApp

No comments yet

Related posts