I’m Pavel, founder of DevGhost. We went through the public Git history of VS Code and Ghost. For several developers, our Git-based estimates show a 5–21× increase in the work reflected in their code changes.

Agent traces turn up in the original commits, replies to reviewers and the fixes made after review:

  • From October 1 to November 14, 2026, 561 of Rob Lourens’ 564 commits say “Written by Copilot”. Example PR.
  • In Ghost, 142 of Hannah Wolfe’s 180 PRs have direct agent traces, including follow-up fixes. Example PR.

Monthly productivity estimates for VS Code contributors, with growth of 15× and 21× from December 2025 to October 2026.

The chart’s 15× and 21× labels compare each contributor’s December 2025 output with their own October 2026 output.

Full analysis, estimation method and PR examples.

If you use coding agents, how does the 5–21× increase we saw compare with your own experience?

  • jubilationtcornpone@sh.itjust.works
    link
    fedilink
    arrow-up
    28
    ·
    2 hours ago

    FFS. I don’t have any issue with using LLM’s for coding but the number of lines committed is not and has never been an indicator of “productivity.”

    Typically, the more lines someone (or something) shells out, the more concerns I have about their quality of work.

    And with the proliferation of vibe coding, more lines is an increasingly common sign that the person who made the PR doesn’t even understand the code they’re trying to merge.

  • endofline@lemmy.ca
    link
    fedilink
    English
    arrow-up
    4
    ·
    1 hour ago

    If your problem is “agile velocity” or typing speed, maybe you shouldn’t code. Code is a liability, the less you have it, the better. Unfortunately, LLM folks seem to have quite opposite stand to this. Just to inform, I’m not against the llms, I’m against mindless usage of it just for the sake of using it

  • PonyOfWar@pawb.social
    link
    fedilink
    arrow-up
    11
    ·
    2 hours ago

    Output is meaningless. I could easily increase my output tenfold by writing code as verbose as possible, but that would make me a worse developer, not a better one. What have they actually achieved with that output? Is the software stable and maintainable?

  • Kissaki@programming.dev
    link
    fedilink
    English
    arrow-up
    3
    ·
    2 hours ago

    Looks like they didn’t check whether the 10x is 9 “fixups” until the tenth change is what would have traditionally been the real solution in one change.

  • RedWeasel@lemmy.world
    link
    fedilink
    English
    arrow-up
    4
    ·
    2 hours ago

    Can and Should be used are drastically different. How many people using these even know what the code they are committing is actually doing or can explain what it is doing. Using AI is fine if you are using it carefully, but alas my gut says most aren’t using it that way.

  • kureta@lemmy.ml
    link
    fedilink
    arrow-up
    2
    ·
    edit-2
    2 hours ago

    If you want to do something that has been done millions of times before, such as data visualization, it helps greatly. If you are making anything remotely novel and complicated it is a great foot-gun unless you already know what you are doing and what you want.