AI Coding Agents: Programming’s Third Era

AI coding agents compared with punch cards and code editors in programming history

AI coding agents mark a third era of programming, and one engineer argues it rivals the jump from punch cards to editors. The claim is that agents are a new paradigm, not better autocomplete, and that people are still programming, only now in English.

What links the three eras is the feedback loop: the time between making a change and learning whether it worked, and who runs that cycle.

Punch cards: days per change

Early programs started on paper. IBM engineer Dale Fisk, who learned the trade around 1973, filled out coding forms by hand, one form line per card. Someone else did the punching: he dropped the sheets in a keypunch operator’s in-basket and collected his deck the following day.

He wasn’t alone. Martin Gorfinkel, who started in 1960, and Janice Silcox, who programmed at Sperry Rand’s UNIVAC division from 1968, both described handing sheets to a keypunch shop. Some shops kept a few machines for quick fixes, and many universities offered self-service rooms.

After punching, the deck waited in a queue. Cornell promised minutes for small jobs, hours for long ones and overnight for very long ones, with low-priority work stretching to five days. Silcox considered a two-hour wait remarkably quick. One typo meant a new card and a full rerun.

Three flat plates show punched holes, horizontal instruction bars, and eight tiles around a central tile.
Image: Ryan X. Charles

Editors: seconds per change

Interactive editing grew out of time-sharing. In 1962 Fernando Corbatรณ and colleagues at MIT described CTSS, prompted by the slowness of debugging in batches. Dan Murphy wrote TECO that year on MIT’s PDP-1. Berkeley’s QED came in the mid-1960s, the first Unix manual listed ed in 1971, and both vi and Emacs trace back to 1976.

The change became normal through the 1970s. Fisk’s lab got shared terminals around 1973, and within a month he had “set aside my punched cards.” Cornell had about 150 terminals by 1978, and student batch work crested that year before declining. Turbo Pascal, released in November 1983, put editor and compiler in one program. Visual Studio Code was announced in April 2015, and 75.9% of developers in Stack Overflow’s 2025 survey said they use it.

The loop shrank from hours to seconds, and the programmer ran every step personally: write, run, read the error, fix.

Three flat plates show punched holes, horizontal instruction bars, and eight tiles around a central tile.
Image: Ryan X. Charles

Agents: the loop moves up

Anthropic launched Claude Code on February 24, 2025 as a limited research preview, describing it as its first agentic coding tool. It could read and search code, edit files, run tests, push commits to GitHub and call command-line tools. General availability followed on May 22, 2025, and OpenAI and Google released terminal agents within months.

Look at that list and you see the editor-era loop, now performed by the agent: it writes code, runs it, reads the error and fixes it.

Adoption figures show the pull. Anthropic’s April 2025 analysis found 79% of Claude Code conversations were automation, versus 49% on Claude.ai. By February 2026 Claude Code’s run-rate revenue passed $2.5 billion, and its weekly users had doubled since January 1.

The human’s loop didn’t vanish. It moved up to intent, review and the workflow that keeps output correct.

Why the medium changed, not the job

Programming has always climbed levels of abstraction, from machine code to assembly to high-level languages. In January 2023, Andrej Karpathy wrote that “the hottest new programming language is English.” Fred Brooks quoted David Parnas decades ago saying automatic programming has always meant programming in a higher-level language than was available. The argument is that agents fit that pattern: an issue with a clear outcome and a way to check it, written in English, works as a program that the agent compiles into code.

There’s an irony here. In the punch-card era many programmers specified a program and someone else produced the artifact. The editor era, when everyone read and wrote every line, may turn out to be the exception.

Objections and the case for this time being different

  • Trust. In Stack Overflow’s 2025 survey, 46% of developers distrust the accuracy of AI tools. The author’s reply is that this decides where effort goes, as tests, review and deterministic workflows take over from card-by-card checking.
  • Adoption. Only 14.1% of developers use AI agents daily at work, while three in four use VS Code. Cards also lingered for years after terminals arrived.
  • Past hype. A 1981 British program generator was called The Last One, and James Martin’s 1982 book was titled Application Development Without Programmers. Later waves of CASE tools, fourth-generation languages and low-code platforms made similar promises. Edsger Dijkstra argued natural-language programming was foolish because formal notation’s precision is a feature.

The author says those tools shrank the problem to fit a generator, form or diagram and failed at the edges. Agents instead use the same general-purpose languages and codebases people do, and they run the edit-run-fix cycle on their own.

What the job becomes

If that holds, reading and writing code stops being the main task. Programmers decide what to build, set constraints, define how success will be checked and improve the workflow that builds software correctly. The hard part shifts from writing code to knowing whether it’s right.

Dijkstra was right that precision matters. It just moves from syntax into specifications and checks. The next data point to watch is whether daily agent use among developers climbs from 14.1% as the 2025 pattern with terminals suggests it could.

Source: Ryan X. Charles

Written by

Saksham Bhargava

Saksham Bhargava runs Alteroid and has been writing about consumer technology for over five years, covering iPhone, Android, Windows, AI tools and gaming. With an MCA and six years of hands-on web development experience, he tests the features he writes about on his own devices and focuses on clear, practical guides that solve real problems. Away from the screen, he enjoys travelling to places most tourists skip.