May 31, 2026
Bots Behaving Badly
Backpressure is all you need
Let the bot work, but only if the machine slaps its hand before your coworkers do
TLDR: The article argues AI coding helpers should be forced to catch more of their own mistakes before humans review anything. Commenters split fast: some said this is obvious common sense, while others mocked the whole trend as overcomplicated robot babysitting.
A fresh coding-advice post tried to split the difference in the great "should we let the AI run wild?" fight, and the comments immediately turned into a mini office civil war. The article’s big idea is simple: don’t let an artificial intelligence helper dump messy work on humans. Instead, make the computer check its own homework first with guardrails like tests, type checks, and other automated blockers so reviewers only see work that’s at least basically sane.
That sounds reasonable, but the crowd was very not united on whether this was genius, obvious, or totally overblown. One camp basically yawned and said, welcome to 101. User dnnddidiej came in with the energy of a fire alarm: “Anyone not doing this? If not do it now!” Another commenter, zoltan, was even more dismissive, arguing this has been obvious for months and name-dropping earlier agent-loop ideas like this was yesterday’s news.
But the skeptics were sharper — and funnier. vermilingua saw the line about reducing junk pull requests and replied with a deadpan “Oh boy,” which is internet shorthand for absolutely not, this is going to be a mess. Meanwhile, wellpast pushed back on the whole grand-plan vibe, saying real creative work happens through tiny, quick back-and-forth steps, not giant automated workflows. In other words: some readers want a disciplined robot employee, others think people are building a full theme park ride just to avoid typing carefully.
Key Points
- •The article says fully unattended coding agents can generate bugs, confusing changes, and too many pull requests for humans to review effectively.
- •It says requiring human review at every small step is safer but reduces the productivity benefits of using an agent.
- •The proposed alternative is to let agents validate more of their own work before human review.
- •The article defines backpressure as a downstream mechanism that limits upstream work and applies this idea to software development workflows.
- •It identifies tests, type systems such as TypeScript, and CI-based guardrails as mechanisms that catch issues earlier and leave humans to review readability, complexity, and design.