What does an IDE actually do?

The full story

The episode's narration, word for word. Numbers link to the sources below; timestamps play that moment.

The question 0:00

All season we've been writing code, and now we look at where people write it. Yet code is just text, and you could write every program in this series in the plainest text editor. In episode one we called the IDE, the integrated development environment, the workshop. So what does that workshop actually do?

The loop 0:23

Programmers spend their days in one loop. Write some code. Run it. See what broke. Fix it, and run it again. Nearly everything an IDE does makes that loop shorter, or catches the problem before you press run. Look at how long it used to be.

A line per card 0:45

In 1928 IBM settled on a punched card with 80 columns, and for decades cards were how programs got into computers ¹. One card held one line of code, so a program was a stack of cards ¹. Computers were too expensive to sit idle, so jobs were collected and run together in batches ². You handed in your deck and waited for the result. One typo meant a new card and another trip around the loop.

One question mark 1:15

From 1963, time-sharing at MIT let people type to a computer through a terminal ². On Unix, the editor was ed, and its 1971 manual calls it "the standard text editor" ³. You edited one line at a time, often on a terminal that printed onto paper ^3,4. The manual lists exactly one error message: a question mark, for any error ³. Richard Stallman later compared editing like this to blindfold chess, because you had to keep the text in your head ⁴.

We ran it 1:49

ed still comes with macOS, so we ran it. We typed a line of Python with a typo, printed it, then typed "print it", in English. ed replied with a question mark, just as in 1971 ³. We fixed it, saved, and Python printed hello. It works, but you never see the whole file, and ed tells you nothing about whether your code is right.

Seeing the board 2:16

Then came screens ⁴. In a display editor, the text sits on the screen and changes as you type, so, in Stallman's words, you can see the board ⁴. Emacs was working at MIT by late 1976 ⁴. Around the same time at Berkeley, Bill Joy wrote vi, for new screen terminals ⁵.

The editor war, fairly 2:39

Programmers still argue over them. In vi, keys mean different things in different modes: one mode for typing, another for commands ⁵. Emacs has no separate typing mode: keys always type, and commands sit on Control and Meta, and you can reprogram almost all of it ^4,5. Both survive: in Stack Overflow's 2025 survey, 24.3 percent of respondents used Vim regularly in the past year, and 14 percent its offshoot Neovim ⁶.

Turbo Pascal 3:11

In 1983, a small company called Borland sold Turbo Pascal ^7,8. Its compiler was written by Anders Hejlsberg, who we met in episode six ⁸. It put an editor, a compiler and a run command into one program that needed as little as 28 kilobytes of disk ^7,9. It cost $49.95 by mail order ^9,8. It wasn't the first integrated environment, but it made one cheap and fast ^8,10.

Integrated 3:41

That's what integrated means. When the compiler hit a mistake, Turbo Pascal opened the right file in the editor and pointed at the error ⁷. Going from a mistake to its fix now took a keypress, not a new deck of cards ⁷. Modern IDEs add five more things, and most of them rest on one trick: the editor reads your code the way the language does.

Colours 4:06

First, colour. Before a language runs code, it splits it into pieces called tokens: keywords like def and return, names you made up, text in quotes, and comments for humans ¹¹. An editor can make the same split as you type, keeping the comments too, and colour each kind ¹¹. It isn't just pretty: forget a closing quote, and in many editors the rest of the line turns the wrong colour.

Mistakes before you run 4:35

Second, catching mistakes before you run. The editor keeps handing your code to a checker that knows the language, and underlines what fails ¹². Here we left off a closing bracket, and Python's checker says it was never closed, on line four. Add it, and the program prints twelve. It's like a spell-checker, with a limit: it checks grammar, not whether the program does what you meant.

Autocomplete 5:02

Third, autocomplete. The editor knows that name holds text, so it can ask what text can do ¹². Type a dot and the letters s t, and Python itself lists two choices: startswith and strip. This kind of autocomplete doesn't guess; it only offers what actually exists.

The debugger 5:24

Fourth, the debugger, for the bugs that checking can't see ¹³. Turbo Pascal's 1984 advert already boasted that it located run-time errors in your source ⁹. By 2015, VS Code's launch page promised you could "set breakpoints, step into and out of your code, and inspect variables" ¹⁴. A breakpoint is a line where the program pauses, so you can look inside ¹³. This function should average four, six and eight, which is six, but it prints four point six six. We paused it with Python's debugger, pdb ¹³. On the first pass, i is already 1, so the 4 in position 0 was skipped. Start counting at zero, and it prints six.

Rename, not find-and-replace 6:14

And refactoring: reshaping code without changing what it does. Say you rename a variable called total. Find and replace changes every "total", including the word inside the quoted label, so the output changes. An IDE's rename changes only the variable, because it knows which "total" is which. We ran both, doing the rename with Python's own tokenizer, and only the rename kept the original output.

Every editor, every language 6:47

But all of this means the editor must understand each language, and that work was repeated for every editor ¹². Five editors and ten languages would need fifty plug-ins. In June 2016, Microsoft, with Red Hat and Codenvy, announced a shared answer: the Language Server Protocol ¹⁵. Each language gets one language server, a separate program that understands it, and any editor can ask it questions in a standard text format called JSON ^15,12. Five editors and ten languages now need fifteen pieces.

Asking a real server 7:24

So we asked a real one. clangd is a language server for C and C++, from the LLVM project ¹⁶. We gave it a tiny C file and sent the message an editor sends when you choose "go to definition" ¹². It answered: line zero, characters four to ten, exactly where square is defined, counting from zero.

Why VS Code spread 7:49

The protocol came out of Microsoft's VS Code, launched in April 2015, free, for Windows, Mac and Linux ^14,15. Its launch page pitched it as working "without the need for a full IDE" ¹⁴. In Stack Overflow's 2025 survey, 75.9 percent of respondents used it regularly in the past year, against 29 percent for Visual Studio ⁶. Whether that makes it an IDE is the same kind of naming argument we met in episode one.

Then came AI 8:22

In June 2021, GitHub previewed Copilot, which suggests whole lines or entire functions as you type ¹⁷. Now there are many such assistants, including Cursor, Claude Code, OpenAI's Codex, Google's Antigravity and Amazon's Kiro ^{18,19,20,21,22}. Claude Code is made by Anthropic, whose Claude helps research and script this channel, so we're naming it, not recommending it ¹⁸. In Stack Overflow's 2025 survey, 84 percent of respondents used AI tools or planned to ²³.

A different kind of help 8:58

Our autocomplete listed only what exists. An AI assistant predicts what code is likely to come next, from patterns in huge amounts of public code ^17,24. Think of a well-read helper who suggests what usually comes next. Some newer assistants can also run your code and tests, but each suggestion is still a prediction ¹⁸. Here the picture breaks: it isn't a person, it doesn't know what you meant, and it can sound sure while being wrong ²⁴.

What the evidence says 9:29

So what can they be trusted with? In a 2022 study, the 2021 Copilot was given 89 coding tasks chosen because they're easy to get wrong on security, and about 40 percent of its 1,689 programs had security weaknesses ²⁴. In a 2023 study of 47 people using a 2022 OpenAI model, those with the assistant wrote less secure code, yet were more sure it was secure ²⁵. And a 2025 study of models from 2023 and 2024 found that, on average, at least 5.2 percent of the software packages commercial models suggested didn't exist ²⁶. Newer tools may do better or worse, so every result carries its date.

Faster, or feeling faster? 10:19

Speed is murkier. In an early 2025 trial by the research group METR, experienced developers using AI tools took 19 percent longer, while believing AI had cut their time by 20 percent ²⁷. In 2026, METR said that result was out of date, and that newer data hinted at a speed-up but was unreliable ^27,28. Capability is rising, though: METR finds that the length of software task AI agents can finish half the time has been doubling every four to seven months ^29,30. More developers distrust these tools' accuracy than trust it ²³.

So what does an IDE do? 11:00

So what does an IDE actually do? It shortens the loop between writing code and finding out it's wrong ⁷. It reads your code the way the language does, so it can colour it, check it, complete it and rename in it safely, and it can pause it as it runs ^11,13. And now it hosts assistants that write plausible code fast, which those same tools help you check ²³. Which programming language should you learn first?

Sources

Every factual claim in the episode is tied to one of these. Spotted an error? Tell us.

1. IBM, "The IBM punched card" (IBM History) — ibm.com
2. IBM, "Time-sharing" (IBM History) — ibm.com
3. Unix Programmer's Manual, First Edition, "ed (I)", 3 November 1971 (cat-v.org… — man.cat-v.org
4. Richard Stallman, "EMACS: The Extensible, Customizable Display Editor", ACM Conference… — gnu.org
5. Jim Joyce, "Interview with Bill Joy", Unix Review, August 1984 (archived transcription) — web.archive.org
6. Stack Overflow Developer Survey 2025, Technology (Dev IDEs) — survey.stackoverflow.co
7. Borland International, "Turbo Pascal Reference Manual", February 1984 (copyright 1983,… — bitsavers.org
8. Wikipedia, "Turbo Pascal" (accessed 7 October 2026) — en.wikipedia.org
9. BYTE, February 1984: Borland advertisement "Turbo Pascal" and Jerry Pournelle, "Chaos… — archive.org
10. Wikipedia, "Integrated development environment" (accessed 7 October 2026) — en.wikipedia.org
11. Python 3 documentation, "tokenize: Tokenizer for Python source" — docs.python.org
12. Microsoft, "Language Server Protocol: Overview" — microsoft.github.io
13. Python 3 documentation, "pdb: The Python Debugger" — docs.python.org
14. Visual Studio Code home page as archived on 29 April 2015 — web.archive.org
15. The VS Code Team, "A Common Protocol for Languages", Visual Studio Code blog, 27 June 2016 — code.visualstudio.com
16. LLVM project, "What is clangd?" — clangd.llvm.org
17. GitHub blog, "Introducing GitHub Copilot: your AI pair programmer", 29 June 2021 — github.blog
18. Anthropic, "Claude Code overview" (documentation) — docs.anthropic.com
19. Cursor, home page (accessed 7 October 2026) — cursor.com
20. OpenAI, "Codex IDE extension" and "Codex CLI" (documentation) — developers.openai.com
21. Google, "Google Antigravity" (home page, accessed 7 October 2026) — antigravity.google
22. Kiro, "Kiro Documentation" — kiro.dev
23. Stack Overflow Developer Survey 2025, AI — survey.stackoverflow.co
24. Hammond Pearce et al., "Asleep at the Keyboard? Assessing the Security of GitHub… — arxiv.org
25. Neil Perry, Megha Srivastava, Deepak Kumar, Dan Boneh, "Do Users Write More Insecure… — arxiv.org
26. Joseph Spracklen et al., "We Have a Package for You! A Comprehensive Analysis of… — arxiv.org
27. Joel Becker, Nate Rush, Elizabeth Barnes, David Rein (METR), "Measuring the Impact of… — metr.org
28. METR, "We are Changing our Developer Productivity Experiment Design", 24 February 2026 — metr.org
29. Thomas Kwa et al. (METR), "Measuring AI Ability to Complete Long Software Tasks",… — arxiv.org
30. METR, "Time Horizon 1.1", 29 January 2026 — metr.org

Researched and scripted with AI assistance, fact-checked claim by claim, with synthetic narration and diagrams drawn in code. How we make episodes.