The compiler was supposed to end programming too

It’s 7:40 on a Tuesday dawn in 2031, and you haven’t written a row of code yet. You likely won’t today, and that’s fine.

While you slept, four agents worked through the backlog you sorted yesterday afternoon. Three of them finished. Their drag requests are waiting for you, all alongside the tests it added, a abbreviated record of the characteristic working, and a note concerning the parts it wasn’t certain of. The fourth delegate stopped at 2:13 a.m. and remaining you a inquiry instead:

The spec says refunds go rear to the first card. This customer’s cardstock expired in March. What should happen?

The delegate could compose refund code all day long. What it can’t do is decide what your business owes a client whose cardstock has expired, so it stopped and asked, which is exactly what you’d desire from a new employ too. Your dawn goes akin this: you peruse three drag requests, you go and conversation to person in support, and you compose the regulation downward so plainly that nobody (human or otherwise) has to conjecture again.

Imagined · Tuesday, 2031

Agents

1234

You

5678910

22:0002:0006:0010:0014:0018:0022:00

Agents

1. 122:00 Three agents activity through yesterday's backlog and open three drag requests, all alongside tests and a recording.
2. 222:00 A fourth delegate stops and leaves a question: refunds go to the first card, but that cardstock expired.
3. 313:00 Agents build the refund regulation you wrote down, alongside the tests you wrote first.
4. 421:30 Tonight's runs commencement from the scheme you approved.

You

1. 507:40 Read the three drag requests and observe the recordings.
2. 609:00 Answer the 02:13 question. You can't, yet.
3. 710:00 Ask assistance what customers alongside expired cards expect.
4. 811:00 Write the regulation downward as a abbreviated spec and five tests.
5. 914:00 Read the agents' activity on the one part that touches money.
6. 1016:30 Approve tonight's plan. Go home.

I made that Tuesday up. None of it is discipline fiction, though. Agents that activity unattended overnight be today, and one of them runs on my own device all night at ten. It invents a small project, builds it, tests it, and I discover out what it made complete breakfast. The archetypal morning, I had no idea what it had built. That build is now Rule Safari, a live project on this site, alongside an article explaining how it works.

So is this the end of programming? People have been asking that for additional than seventy years. I discover that comforting. I additionally discover it a small suspicious.

Programming has been ending since 1954

On 10 November 1954, IBM’s Programming Research Group completed a preliminary study on a new tongue called FORTRAN. One declaration in it says that FORTRAN “should practically eliminate coding and debugging.”

I affection this sentence. It’s so certain of itself, and it equal came alongside a budget: problems solved “for small than fractional the cost” of doing it without FORTRAN. The industry had a name for this benevolent of item rear then. Assemblers and compilers were called automatic programming. You’d depict what you wanted, and the device would compose the program.

FORTRAN shipped in 1957, and it really did eliminate a benevolent of coding. Almost nobody translates formulas into device instructions by hand anymore. Debugging survived fair fine. (This duration I chased a Scry bug that came downward to a bookmark saved lone all dozen hours, so I can verify debugging is living and well.)

What FORTRAN really did was obtain one tedious part of the job distant and move the job up a level. People stopped thinking concerning registers and started thinking concerning bigger problems. Programs got cheaper to write, so everybody wanted additional of them. That has happened complete and complete since.

1. 1947

   Assembly languageYou stopped: penning instructions as raw numbersKathleen Booth, "Coding for A.R.C."

2. 1957

   FORTRANYou stopped: translating formulas into device instructions by handThe 1954 report: it "should practically eliminate coding and debugging."

3. 1972

   CYou stopped: rewriting the identical program for all new machine

4. 1981

   The Last OneYou stopped: nothing, as it turned outA BASIC program generator, named as the final program you would always need to buy.

5. 2008

   Stack OverflowYou stopped: solving all error communication alone

6. 2010

   npmYou stopped: penning your own everything

7. 2021

   GitHub CopilotYou stopped: typing the boilerplate

8. 2025

   Coding agentsYou stopped: typing most of the code

9. 20??

   NextYou stopped: ?

My favorite rung is 1981. A British business called D.J. “AI” Systems (yes, really) sold a BASIC program generator named The Last One. The name was the entire pitch. It would compose all your programs for you, so it would be the final program you’d always need to buy.

Reviewers established it helpful for several endeavor programs. It was not the final one.

Here’s the form on all rung. Something gets automated, the group doing the activity halt doing that item by hand, and the site keeps expanding anyway. GitHub counted additional than 180 myriad developers in its 2025 report, and additional than 36 myriad of them had joined in the former dozen months.

So I don’t think AI ends programming. I do think this rung is distinct in one way, and that’s speed. GitHub Copilot’s preview came out in June 2021. By 2025, 84% of developers in Stack Overflow’s study were using AI tools or preparedness to. Four years, from novelty to default.

What 2026 really looks like

The vibes are noisy correct now, so let’s commencement alongside numbers. (Speaking of vibes: Andrej Karpathy coined “vibe coding” in February 2025, for the benevolent of coding anywhere you “fully provision in to the vibes.” Stack Overflow asked concerning it that year. 72% of developers stated it isn’t part of their expert work, and another 5% stated emphatically no. I affection that they offered that option.)

Stack Overflow’s study is the biggest regular check-up we have, and the acceptance curve in it is steep: from 70% of developers using AI tools or preparedness to in 2023, to 84% in 2025, whenever fractional of expert developers were using them all day.

You’d think all that use would average developers affection these tools. Nope. Favorable opinion cut from 77% in 2023 to 60% in 2025, and lone ticked rear up to 62% this year. In 2025, additional developers distrusted the accuracy of AI output (46%) than trusted it (33%). The top frustration, picked by 66% of them, was “AI solutions that are nearly right, but not quite.”

I awareness seen. Almost correct is an costly benevolent of wrong, since you have to peruse all row to discover the part that isn’t. (45% of them stated debugging AI-generated code takes longer.)

Then there’s my favorite study of the entire AI boom, mostly since it’s so rude. In first 2025, METR, a nonprofit investigation group, ran a appropriate randomized controlled trial. Sixteen informed open-source developers worked on 246 genuine issues in their own repositories, projects they’d been on for years, and all matter was randomly assigned to authorize AI tools or ban them.

Before they started, the developers expected AI to cut their period by 24%. When they finished, they believed it had cut it by 20%. The clock stated that alongside AI they took 19% longer.

Nineteen percent longer! And they walked distant certain it had helped.

← longershorter →

• Economists predictedexpert forecast39% small time
• ML researchers predictedexpert forecast38% small time
• The developers expectedbefore the tasks24% small time
• The developers feltafter the tasks20% small time
• The clock measuredwhat happened19% additional time

Before anyone screenshots that for a hot take: it’s one study of sixteen people, using early-2025 tools, on codebases they knew improved than any example could.

And METR tried to run it again. A bigger study started in August 2025, alongside 57 developers and additional than 800 tasks, and in February 2026 METR stated the data was too unreliable to lean on. Mostly that was since developers kept declining to obtain part, since they “do not desire to activity without AI.” The first numbers hinted at an 18% speedup for the returning developers (with a range broad adequate to contain no speedup at all).

I think that’s the most 2026 way an test has always failed. People refused to be in the collection without AI.

What I obtain from METR is narrower than any headline, and additional useful: how accelerated you awareness is not a measurement. If you desire to cognize whether AI makes your squad faster, measure it the way you’d measure item else.

The row that keeps going up

METR’s another big project tracks how lengthy a project an AI delegate can complete on its own. They period individual experts on a suite of application tasks, afterward discover the project dimension anywhere an delegate succeeds fractional the time. In 2019, GPT-2 managed tasks that obtain a individual concerning three seconds. GPT-4 got to concerning four minutes in 2023. Claude 3.7 Sonnet reached an hr in February 2025, and Claude Opus 4.6 reached concerning dozen hours a twelvemonth later.

Across the entire series, that figure has doubled concerning all six months. Since 2023, it’s been nearer to all four months.

Now, a log-scale diagram alongside a dashed row on the end is how a lot of bad predictions start, so let me be careful. “Half the time” is doing a lot of activity in that definition. Benchmark tasks are cleaner than your codebase. And METR stopped updating that leaf in September. But if the row keeps going equal roughly, agents volition grip a duration of individual work, fractional the time, about 2028. That’s anywhere my 2031 Tuesday comes from. I didn’t drag it out of lean air. I pulled it out of a log-scale chart, which is lean air alongside axes.

The companies construction these things are already living a type of it. Google stated in October 2024 that additional than a fourth of its new code was generated by AI, and “well complete 30%” six months later. In March 2025, Anthropic’s Dario Amodei predicted AI would be penning 90% of code inside three to six months. That was too soon. But by May 2026, Anthropic stated additional than 80% of the code merged into its own codebase was written by Claude.

When code gets cheap, the queue moves

Make one stage of a procedure faster and the activity piles up in forefront of the next slowest step. Anyone who has worked a galley on a Saturday night knows this: speed up the grill and the plates commencement waiting at the pass.

That’s what’s happening to application teams. In 2025, Faros AI looked at telemetry from additional than 10,000 developers on 1,255 teams. Developers on teams alongside elevated AI acceptance completed 21% additional tasks and merged 98% additional drag requests, and their assessment period went up 91%.

Faros ran it again in 2026 alongside 22,000 developers. Task completion per developer was up 34%, which appears awesome until you peruse the rest. Median assessment period had grown five times. 31% additional drag requests were merging without any assessment at all, and the proportion of incidents to drag requests had additional than tripled.

The code got cheaper, the study got additional expensive, and several teams quietly stopped reading. That matters since the code isn’t continually good. Veracode tested additional than 100 tongue models on 80 coding tasks in 2025, and 45% of the results had a safety flaw. GitClear established that copy-pasted lines went from 8.3% of changed code in 2021 to 12.3% in 2024, during moved code, the benevolent you get from refactoring, cut from concerning a fourth to under 10%.

Spare a idea for the maintainers, too. Daniel Stenberg, who leads curl, started banning group who submitted “AI slop” bug reports in May 2025, penning that the project was “effectively being DDoSed.” In January 2026, curl close downward its bug bounty.

Decide what to build

Write it down

Write the code

Review it

Check it works

Ship and run it

Today: code arrives faster than group can peruse it, so drag requests delay for review.

Fred Brooks saw a type of this coming in 1986. In “No Silver Bullet” he divided the difficulties of application into two piles, borrowing Aristotle’s words: regulation and accidents. The accidents are the difficulty our tools logic us, akin syntax, boilerplate and fighting the build. The regulation is what he called “the specification, design, and evaluation of this conceptual construct,” which is a fancy way of saying: operating out what the item should do. AI is getting extremely fine at the accidents. The regulation is the refund inquiry from 2:13 a.m., and it’s motionless on your desk.

The rung we power be sawing off

This is the part of the data I akin least.

In August 2025, Erik Brynjolfsson and two colleagues at Stanford’s Digital Economy Lab used ADP payroll data to appearance at who was getting hired. Employment for application developers old 22 to 25 was downward nearly 20% from its highest in delayed 2022. Older developers showed no noticeable change.

They updated the study this August, alongside data through June 2026. Young workers in AI-exposed jobs are now 19% below anywhere they’d be if they had kept gait alongside less-exposed peers. Experienced workers display no comparable gap. Most of the consequence comes from companies hiring small young group in the archetypal place.

The remainder of the market isn’t exactly roaring either. Indeed’s indicator of application betterment job postings sits concerning 21% below February 2020, following bottoming out 39% below in May 2025.

The Stanford authors are careful to call their findings descriptive, not causal, and plentifulness alternatively has happened since 2022. Interest rates went up, and a lot of companies had hired too many group during the epidemic (Indeed’s indicator peaked at additional than twice its 2020 flat in first 2022). But the form lines up uncomfortably fine alongside what agents are fine at: the small, well-defined ticket we used to provision the new individual on the team.

Senior engineers get made on the job, mostly by being juniors on genuine teams: doing the boring tickets, breaking manufacturing once, and having person benevolent explain why. If agents obtain all the boring tickets, anywhere does the next elder engineer arrive from? It’s difficult to assessment code fine if you’ve never written much of it.

I don’t have a spotless answer. My conjecture is that the teams doing fine in 2035 volition be the ones that kept hiring juniors in 2026 and gave them a distinct apprenticeship: reviewing delegate activity alongside a elder beside them, owning one small scheme end to end, and penning the specs and tests the agents build from. Writing those specs and tests is a genuine job. It power rotate out to be the job.

Five guesses for 2031

Okay, now the guessing part. Bookmark this leaf and arrive rear in 2031 to chuckle at me.

Teams get smaller, and there are a lot additional of them. According to Y Combinator’s Jared Friedman, a fourth of the Winter 2025 lot had codebases that were concerning 95% AI-generated. In February, Block cut concerning 4,000 jobs, and Jack Dorsey pointed to AI tools and “smaller and flatter teams.” I could be incorrect here, but I motionless don’t think that adds up to small application jobs overall, and neither does the US Bureau of Labor Statistics, which projects 10% growth for application developers, QA analysts and testers from 2025 to 2035.

When item gets cheaper, we lean to desire additional of it. Economists call this the Jevons paradox, following William Stanley Jevons, who asserted in 1865 that additional productive steam engines would create Britain scorch additional coal. Software that took a squad a fourth volition obtain one individual a week, and a lot of problems nobody could oversee to resolve volition abruptly be value solving.

The spec and the tests rotate into the source. Today we treat the code as the item we own and the spec as the document nobody reads. I think that flips. The spec and the tests rotate into what you assessment and contend about, and the code becomes additional akin a compiler’s output: you can open it and inspect it, but you rarely edit it by hand. (Yes, that’s the 1954 commitment again. Maybe this period it sticks.)

Software for an spectators of one. My nightly delegate builds things nobody asked for, including me, and several of them are delightful. When a operating tool expenses an evening alternatively of a quarter, group volition create application the way they create spreadsheets today: for one issue and one person, thrown distant whenever it’s done.

Reading code becomes a career. We have entire careers built about penning code and nearly none built about study it. I anticipate that to change, alongside item akin “verification engineer” on job boards and much improved tooling that shows a reviewer the recording, the risky lines and the tests the delegate didn’t write. The Faros numbers say we need it now, never intellect in 2031.

Every alter carries a label. I think a drag petition volition routinely say who or what wrote it, who checked it, and what was really verified. I already do a small type of this: all project my nightly delegate builds says so, in its README and on this site.

Back to Tuesday

FORTRAN did eliminate a benevolent of coding. It didn’t eliminate coders. It made a lot additional of them, operating on harder problems, and the debugging came alongside for the ride.

I think AI does the identical thing, faster and messier. The developer of 2031 spends the day deciding what should exist, explaining it plainly adequate that a device can build it, and checking that what came rear is right. That was continually the difficult part. We fair had small period for it since we were occupied typing.

What does your Tuesday appearance akin correct now? Are you reviewing additional than you write? Tell me in the comments. I’d akin to cognize if my conjecture is already wrong.

Written by hrih