{"slug": "the-ultimate-equation-about-ai-coding", "title": "The Ultimate Equation about AI Coding", "summary": "Software developer and blogger Jason Gorman argues that LLM-assisted and agentic code generation has raised output but not user engagement or the bottom line, because faster shipping degrades the feedback loop that produces real product value. Gorman cites multiple large-scale studies finding LLMs are good at generating code but bad at reliably modifying it, and warns that optimizing only one part of the feedback loop amounts to software engineering's \"original sin.\" He concludes that teams should slow workflows, take smaller steps and keep humans closer in the loop rather than chase fully autonomous agentic horizons.", "body_md": "There are some useful mathematical models relating to LLM-assisted and agentic software development, like my formula for the [reliability of model interactions](https://codemanship.wordpress.com/2026/08/27/agentic-horizons-when-the-wheels-start-to-wobble/) (“steps”) in an agentic workflow. It predicts when the wheels may start to wobble in autonomous execution and therefore how far apart human checkpoints – Actual Intelligence – may need to be to stabilise reliability and keep software shippable.\n\nBut the question that will probably determine the future of this technology is whether or not, in the final reckoning, it’s actually worth it. Is the additional value created greater than the additional cost? Or is the same value being created at lower cost?\n\nAI ROI = net value created with AI – net value created without it\n\nSo far, the best available evidence clearly shows that – while output is undeniably up – that isn’t making a noticeable dent in user engagement or the bottom line.\n\nShipping apps in less time and at lower cost doesn’t save you anything if nobody’s using them. Probably correctly, many people have identified that shipping wasn’t the main problem in the first place. Figuring out what has value arguably always was (and always will be). But, in iterative development, shipping is a major part of figuring that out. *That first release isn’t the finish line – it’s the starting pistol.*\n\nThat’s when our ideas – our guesses and assumptions about what’s needed and what folks will use – get tested in the real world. That’s when we start *learning*. We ship. We learn. We adapt. Ship. Learn. Adapt. Ship. Learn. Adapt. It’s the “Learn. Adapt.” part that turns the base metals into gold.\n\nIt turns out learning is where most of the value gets *discovered*. It’s not usually in what we planned. It’s the feedback loop between what we ship, how reality responds to it, and how we adapt to it that does the heavy lifting.\n\nThis is where LLM-assisted and agentic code generation can – and demonstrably does in most cases – hurt us.\n\nIt hampers the feedback loop that really matters in two ways:\n\n- Shipping faster than reality can absorb tends to produce lower signal-to-noise in the feedback. It increases feedback latency, and actively [*slows* learning](https://codemanship.wordpress.com/2026/07/15/feedback-latency-learning-productivity/) .\n- Iterating software designs to adapt to feedback, it turns out, is not something the technology’s good at by itself. In fact, [multiple large-scale studies found that it’s really rather bad at it](https://arxiv.org/abs/2603.13428) . LLMs are very good at generating code they’re bad at modifying reliably – if you let that code slip through the net into future model interactions.\n\nThe risk is that we commit what I call software engineering’s “original sin” – we optimise only one part of the feedback loop, at the expense of the whole.\n\nThese outcomes aren’t inevitable. Nobody’s forcing you to ship 100 changes a day. Nobody’s stopping you doing code reviews more often on smaller batches of changes, or refactoring continuously. There’s no police force that will storm the building if you look at the code for problems your quality gates didn’t anticipate.\n\nYes, I appreciate you may lose your LinkedIn bragging rights if you slow the workflow down, take smaller steps and stay closer in the loop – keeping both hands on the wheel. But I’ve yet to see any credible evidence of anyone genuinely producing software of any appreciable complexity – that gets used for real, and is what most users would consider to be reliable enough – who doesn’t.\n\nBut even if we could extend autonomous agentic horizons, we’d walk straight into the feedback latency trap – shipping faster than reality can meaningfully respond to. That’s the thing about reality – it moves at its own pace. As powerful as LLMs and agents are, they can’t make Tuesday come on Monday.\n\nI’m betting that the value equation has a [sweet spot](https://codemanship.wordpress.com/2026/08/28/the-wall-confronting-reliable-coding-agent-autonomy/), just like it does for human software development. And that sweet spot – where the gains from AI-assisted coding outweigh the downstream losses – falls, I suspect, far short of the aspirations of autonomous long-horizon execution. That remains science fantasy – a tale of a highly improbable tomorrow.\n\nMeanwhile, as the media and the markets become more skeptical of the hype and as prices start to reflect the real costs – which they must at some point, with > $1 trillion of debt already accrued – the CFO has entered the chat and she isn’t interested in tales of tomorrow.\n\nA founder or an investor may be sated by being told the gold’s over the next hill. But even the biggest businesses can only function for so long on promises, and we’ve had four years of this agentic alchemy to produce some of that promised gold. And all the indications are that the promises are still a long way off, if they’re possible at all.\n\nTime to put the sci-fi back on the shelf and start looking for a sweet spot in the present, perhaps?", "url": "https://wpnews.pro/news/the-ultimate-equation-about-ai-coding", "canonical_source": "https://codemanship.wordpress.com/2026/09/03/the-ultimate-equation-about-ai-coding/", "published_at": "2026-09-03 05:32:49+00:00", "updated_at": "2026-09-17 07:24:40.496962+00:00", "lang": "en", "topics": ["ai-agents", "ai-tools", "ai-research"], "entities": ["Jason Gorman", "LinkedIn"], "alternates": {"html": "https://wpnews.pro/news/the-ultimate-equation-about-ai-coding", "markdown": "https://wpnews.pro/news/the-ultimate-equation-about-ai-coding.md", "text": "https://wpnews.pro/news/the-ultimate-equation-about-ai-coding.txt", "jsonld": "https://wpnews.pro/news/the-ultimate-equation-about-ai-coding.jsonld"}}