BRENNEN

Industries / Technology / Software and SaaS

Writing code

Turning what a program should do into instructions a computer can run, done by people and now also by AI tools.

7

Milestones achieved, from 1957 to 2025.

2

To watch for: announced, not yet confirmed outside.

2

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1957

    Formulas instead of machine code

    IBM released FORTRAN, a programming language that let scientists and engineers write their problems as formulas, together with a compiler, a program that translates that language into the computer's own instructions, for its 704 computer. IBM says a task that needed as many as 1,000 hand-keyed machine instructions could be written in 47 FORTRAN statements, and that the compiled programs ran nearly as fast as hand-written ones. A FORTRAN program could run on any computer that had a FORTRAN compiler.

    ibm.com

    Credit

    John Backus and his team at IBM

  2. 1968

    Software engineering named

    The NATO Science Committee held a working conference in Garmisch, Germany, from 7 to 11 October 1968. More than fifty people from eleven countries attended. The organizers chose the phrase "software engineering" on purpose, as a provocation: it implied that making software should rest on the theory and discipline of the established branches of engineering. Participants debated what some of them called the "software crisis", a widening gap, in one position paper's words, between promises to users and the performance achieved, and between estimated and actual costs.

    homepages.cs.ncl.ac.uk

    Credit

    NATO Science Committee; conference chairman F. L. Bauer; report editors Peter Naur and Brian Randell

  3. 1975

    Late projects get later

    IBM put Fred Brooks in charge of OS/360, the operating system for its System/360 computers, because the project was running late. IEEE Spectrum calls OS/360 the world's largest software project at the time. In 1975 Brooks published The Mythical Man-Month, essays on what the project taught him. Its best-known line is "Adding manpower to a late software project makes it later."

    LessonMore people do not rescue a late software project, and can delay it further.

    spectrum.ieee.orgtheregister.com

    Credit

    Frederick P. Brooks Jr

  4. 1983

    Code anyone may share

    On 27 September 1983 Richard Stallman announced GNU, a complete Unix-compatible set of programs that he would "give it away free to everyone who can use it." Free and open-source software, code whose license lets anyone use, change and share it, became a base layer of the industry. A 2024 Harvard Business School working paper put its value to the firms that use it at $8.8 trillion, and estimated that firms would spend 3.5 times more on software if it did not exist.

    gnu.orghbs.edu

    Credit

    Richard Stallman and the GNU Project; Manuel Hoffmann, Frank Nagle and Yanuo Zhou for the estimate

  5. 2005

    Git for the Linux kernel

    The Linux kernel's developers tracked changes to their code with BitKeeper, a commercial tool they used free of charge. In 2005 their relationship with BitKeeper's maker broke down and the free use ended. Linus Torvalds and the kernel developers wrote their own tool, Git, a distributed version control system: every developer holds a full copy of the project and its history. Its goals included speed and support for thousands of parallel branches of work. In Stack Overflow's 2022 developer survey, 93.87% of 71,379 respondents used Git.

    git-scm.comsurvey.stackoverflow.co

    Credit

    Linus Torvalds and the Linux kernel community

  6. 2021

    AI suggests the next lines

    On 29 June 2021 GitHub opened a technical preview of Copilot, which suggests "whole lines or entire functions" as a programmer types. It ran on Codex, an AI model built by OpenAI. Such tools are now routine. In Stack Overflow's 2025 survey, 84% of respondents used or planned to use AI tools in their work, and 51% of professional developers used them daily. More respondents distrusted the tools' accuracy (46%) than trusted it (33%).

    github.blogsurvey.stackoverflow.co

    Credit

    GitHub and OpenAI

  7. 2025

    Slower, but felt faster

    METR, a research group that studies AI capabilities, ran a randomized trial. Sixteen experienced open-source developers worked on 246 real issues in their own projects, and each issue was randomly assigned to allow or forbid AI tools, mostly Cursor with Claude 3.5 and 3.7 Sonnet. With AI allowed, issues took 19% longer. The developers had expected AI to make them 24% faster, and afterwards still believed it had made them 20% faster. A follow-up with 57 developers using late-2025 tools, reported in February 2026, pointed to a speedup, but METR called it "only very weak evidence" because many developers would no longer work without AI.

    LessonHow fast a team feels with a new tool is not a measure of how fast it is; that takes a comparison against work done without it.

    metr.orgmetr.org

    Credit

    METR

Watch for

  1. Three quarters of new code

    Sundar Pichai, Google's chief executive, said that 75% of all new code at Google is now generated by AI and approved by engineers, up from 50% the previous fall. He also said that a complex code migration done by agents and engineers together finished six times faster than engineers alone could have done it a year earlier.

    Achieved whenan outside party audits how Google counts AI-generated code, or another company reports a similar share under a published definition that others can check.

    blog.google

    Claimed by

    Google, April 2026

  2. Written only by agents

    AWS released AWS Emulator, a stand-in for AWS cloud services that runs on a developer's own machine, for building and testing applications. AWS says it "was written entirely by autonomous agents", and that its team has not written or reviewed any of the code. Agents checked their work against what AWS calls tens of thousands of fidelity tests, which compare the emulator's answers with those of the real AWS cloud (see the Note "Code no person read").

    Achieved whenoutside users or researchers run their own tests against both the emulator and the real cloud and publish how often the two agree.

    aboutamazon.com

    Claimed by

    AWS, November 2026

Open

  1. 01

    Brooks's law, tested on agents

    A public experiment that gives the same coding tasks to teams of 1, 2, 4, 8 and 16 AI coding agents working in parallel on one codebase. It records how long each team takes, what it costs, and whether the result passes the project's tests. A second condition adds agents halfway through a task, the case Brooks wrote about.

    The brief
    Credit · unclaimed
  2. 02

    Run your own AI trial

    An open-source kit that lets any software team run its own version of METR's trial. It takes the team's planned tasks, marks each one at random "AI allowed" or "AI not allowed", reads start and finish times from the issue tracker and git, and reports the estimated speedup or slowdown with its error bars, in plain language.

    The brief
    Credit · unclaimed

Open 01 · the brief

Brooks's law, tested on agents

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public experiment that gives the same coding tasks to teams of 1, 2, 4, 8 and 16 AI coding agents working in parallel on one codebase. It records how long each team takes, what it costs, and whether the result passes the project's tests. A second condition adds agents halfway through a task, the case Brooks wrote about.

Why it is still open

Published studies compare two agents with one, or parallel work with sequential work. CooperBench (January 2026) found that two agents sharing a codebase succeeded about 30% less often than one agent doing both tasks. CodeCRDT (October 2025) ran 6 tasks and found results ranging from a 21.1% speedup to a 39.4% slowdown. The closest study, by Giuseppe Destefanis and Tomaso Aste (August 2026), ran 1,902 team runs and measured how messages and tokens grow with team size, not how long the work took. No study found measures time, cost and pass rate across team sizes on the same real tasks, or adds agents partway through.

arxiv.orgarxiv.orgarxiv.org

Built with what exists

coding agents that run from the command line, such as Claude Code or GitHub Copilot's agent (see the Note "Effort, not model"); the Cursor Agent SDK to script teams of agents (see the Note "Findings that cite notes"); git worktrees, which give each agent its own copy of the files; tasks and tests from CooperBench.

  1. First step

    Run 10 CooperBench tasks with one agent and with two, and check that your two-agent result lands near the published 30% drop.

  2. Done looks like

    50 tasks run at five team sizes, five times each (1,250 runs), with pass rate, wall-clock time and cost published for every run. A second group reruns 10 tasks and gets results inside the published error bars.

Open 02 · the brief

Run your own AI trial

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

An open-source kit that lets any software team run its own version of METR's trial. It takes the team's planned tasks, marks each one at random "AI allowed" or "AI not allowed", reads start and finish times from the issue tracker and git, and reports the estimated speedup or slowdown with its error bars, in plain language.

Why it is still open

METR's February 2026 update says its own design broke down: developers withheld tasks they did not want to do without AI, and time tracking failed when one person ran several agents at once. A search found no open kit that lets teams run such a trial on their own work.

metr.org

Built with what exists

Python or R for the random assignment and the statistics; the GitHub or GitLab API for issue and commit times; METR's published study design as the protocol.

  1. First step

    Write the assignment and timing script against one public repository's issue history, then test it on 20 of your own tasks.

  2. Done looks like

    Five teams at different companies run the kit for eight weeks each and publish their estimates with confidence intervals. One team's raw data is reanalyzed by an outside party, which gets the same estimate.