June 23, 2026
Loop there it is
The Coming Loop
Programmers are fighting over AI that writes code while they watch
TLDR: Armin Ronacher says the next wave of AI coding is about systems that keep the bot working in repeated rounds, not just one-off prompts, but he worries the results are harder to understand and trust. In the comments, people split between hype, burnout, and blunt skepticism, with many asking whether any of this is actually making life or business better.
A thoughtful blog post about the future of AI coding somehow turned into a very human comment-section therapy session. Writer Armin Ronacher says the big shift is no longer simply asking an AI helper for code, but building “loops” around it so the machine keeps trying, checking itself, and trying again until something finally sticks. In plain English: people are setting up systems where the bot keeps working even after it would normally say, “I’m done.” Ronacher’s big worry is that this can produce code that works on the surface but becomes messy, overcomplicated, and hard for humans to truly understand.
And wow, the community had feelings. One camp basically yelled, “Show me the money first!” with one commenter mocking the whole trend as “high level wanking” until it creates a billion-dollar one-person company or serious profits. Another group sounded more existential than excited, with one developer admitting they feel uneasy and don’t even enjoy work made this way, calling the whole thing a “rat race.” That’s the real drama here: not whether the tools are clever, but whether people actually want to live like this.
Still, not everyone was doomposting. Some commenters said they relate to Ronacher’s need to understand what they ship, especially for serious work, but are happy to let AI loose on side projects. Others leaned into the chaos and dropped links to their own experimental AI loop setups like proud garage inventors showing off a homemade robot. The vibe was half future-of-work debate, half group chat panic, with a little “look at my weird build” energy sprinkled in for comic relief.
Key Points
- •The article describes a growing software pattern in which external harnesses keep coding-agent tasks running beyond a model’s initial completion signal.
- •It distinguishes between the internal tool-use loop inside coding agents and the outer harness-level loop that continues or reroutes work.
- •Armin Ronacher says he has not had much success using hands-off looping workflows for code he deeply cares about because he wants high standards and direct comprehension of shipped systems.
- •The article argues that current model-generated code in such workflows often becomes overly defensive, complex, locally optimized, and weak on strong invariants.
- •Ronacher cites Andrej Karpathy’s comment about models being 'mortally terrified of exceptions' to illustrate why LLMs often handle malformed cases locally instead of preventing them by design.