{"slug": "limited-output-is-a-feature", "title": "Limited output is a feature", "summary": "Johanna Larsson argues that the push to maximize LLM agent productivity in software development is a trap, as startups win by thinking differently and producing value, not by outproducing incumbents with more code. She cites Linear's success from focusing on fewer features and quiet contemplation, urging founders to resist automating everything and instead spend more time thinking.", "body_md": "# Limited output is a feature\n\naka the LLM agent productivity boost is a trap\n\nAs the tech startup scene continues to find more and more advanced ways to use LLM agents to build software, complete tasks, do research and marketing, and anything in-between, I’m wondering if we’re losing track of what matters.\n\nWhile people are posting about whether or [not you still need to do (or should do) code review](https://open.substack.com/pub/charitydotwtf/p/ai-enthusiasts-are-in-a-race-against?selection=d6c54345-306a-42be-b631-99b4005b86fb&utm_campaign=post-share-selection&utm_medium=web&aspectRatio=instagram&textColor=%23ffffff&bgImage=true), or [how they’re building their software factories](https://factory.strongdm.ai/), I’m starting to struggle to see how it relates to actually accomplishing the goals of those startups. Are startups that build and obsess over software factories more likely to create a product that understands its users? Does removing human review from PRs increase the quality of the user experience of the product? Or have we gotten lost in a frantic race to optimize for “AI-enablement” over meaningful features?\n\nWith LLMs you can build pipelines from customer feedback to automated PR creation and approval, to shipping whole features without human intervention. You can parallelize your attention across ten different tasks at the same time. The bottleneck is your ability to come up with things to do, with the hungry machine always looking for more inputs.\n\nBut is that good?\n\n# How to build good product\n\nLinear has historically been a great example of a company that found success by [ not building features](https://www.lennysnewsletter.com/p/linears-secret-to-building-beloved-b2b-products-nan-yu). They created a well-loved product by focusing on the core experience and avoiding adding cruft, by being opinionated about what features to build and which ones to ignore, even when users are asking for them. This type of product is the result not of the lines of code or the number of PRs merged, but “quiet contemplation”. I can’t help but imagine the founders deep in the dark Finnish forests spending their time in the sauna and snow, thinking, only briefly emerging once a week to spend an hour typing out a new feature, before retreating back.\n\nSuccess has never been about [how many lines of code you can produce](https://xcancel.com/HarryStebbings/status/2060736248004710452), especially not for startups. Because lets face it, [a startup can’t beat a large incumbent on that metric](https://xcancel.com/karrisaarinen/status/2060977012471370054). Maybe that’s why software factories are so attractive to early stage startups, because of the frustration of not being able to outproduce the large enterprise companies they’re trying to disrupt.\n\nBut that’s missing the point. Startups don’t disrupt incumbents by outproducing them, they disrupt incumbents by *thinking differently*. Linear didn’t take on Atlassian’s JIRA by having more features than JIRA, they did it by cutting away all of the things that didn’t really matter, revealing the core of the product that had real value. Combining that with a deliberate and curated user experience, they found their product marketing fit by having *fewer* features.\n\n# Think more, do less\n\nSpend more time thinking and less time doing. Stop giving into the temptation of endlessly feeding the machine with increasingly pointless tasks. Don’t fall into the trap of trying to add every single feature someone asks for. Get away from the computer and spend some time thinking. Win by being different, by producing more value, not more code. Resist that urge to hand everything over to the machine, and try the slower more thoughtful approach.\n\nYour users will thank you.\n\nWritten by [Johanna Larsson](/about).\nThoughts on this post? Find me on Bluesky at\n[\n@jola.dev\n](https://bsky.app/profile/jola.dev)\n.\n\n## Related posts\n\n[\nelixir\ncluster\n](/posts/distributed-ratelimiter-with-hrw)\n\n### Distributed rate limiter with HRW in Elixir\n\nBuilding the classic distributed rate limiter, using the HRW library instead of the traditional ExHashRing.\n\n[\nllm\nphilosophy\n](/posts/a-computer-can-never-be-held-accountable)\n\n### A computer can never be held accountable\n\nFor better or for worse the things you create with LLMs are your responsibility, just like with any other tool.\n\n[\nelixir\ncluster\n](/posts/elixir-cluster-101)\n\n### Elixir Cluster 101\n\nA practical guide to clustering your nodes and tracking the cluster state for real use cases.", "url": "https://wpnews.pro/news/limited-output-is-a-feature", "canonical_source": "https://jola.dev/posts/limited-output-is-a-feature", "published_at": "2026-08-04 00:00:00+00:00", "updated_at": "2026-08-04 12:32:20.267845+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-agents", "ai-products"], "entities": ["Johanna Larsson", "Linear", "Atlassian", "JIRA"], "alternates": {"html": "https://wpnews.pro/news/limited-output-is-a-feature", "markdown": "https://wpnews.pro/news/limited-output-is-a-feature.md", "text": "https://wpnews.pro/news/limited-output-is-a-feature.txt", "jsonld": "https://wpnews.pro/news/limited-output-is-a-feature.jsonld"}}