A deliberately provocative post on Hacker News is asking a question that many developers are thinking but few are articulating this bluntly: what useful work is actually being done by the ecosystem of AI agents, harnesses, and subsidized model access that has sprung up over the past two years?
The skepticism underneath the hype
The post acknowledges that large language models are modestly useful as coding assistants. That is a common and reasonable position. Ask a model to explain a function, generate boilerplate, or suggest a fix for an error, and it often helps. The post does not dispute that. What it questions is the layer built on top of that basic utility: the agent frameworks, the orchestration platforms, the million-dollar compute subsidies sold at pennies on the dollar, and the entire ecosystem of tools that promise to turn language models into autonomous software developers.
The poster sees worse software being built at higher prices, astronomical technical debt accumulating in projects that adopt these tools without understanding their limitations, and a thriving ecosystem of vendors cashing in before the inevitable correction. The question is not whether AI is useful. The question is whether the current trajectory of investment and adoption is producing value or producing hype.
What people actually find valuable
The honest answer, stripped of vendor marketing, is that the most valuable uses of AI in software development are the simplest ones. Code completion that saves keystrokes on repetitive patterns. Explaining unfamiliar codebases by summarizing what a function does or how a module is structured. Generating test cases from function signatures. Translating between languages or frameworks when you already understand the logic. These are narrow, well-scoped tasks where the model's ability to predict plausible output aligns with what the developer actually needs.
The less valuable uses, and the ones that generate the most hype, are the ones that attempt to replace developer judgment entirely. Autonomous agents that plan, execute, and verify multi-step tasks. Code generation from natural language specifications without human review. "Vibe coding" where the developer describes what they want and the model builds it. These approaches produce code that often looks correct but carries hidden assumptions, missed edge cases, and structural problems that surface later in production.
The technical debt problem
The technical debt concern is real. When a model generates code that a developer does not fully understand, that code becomes a liability the moment it needs to be modified. The developer cannot reason about it, cannot predict how changes will affect it, and cannot fix it without generating new code through the same process. Each AI-generated function that passes review without being understood creates a small pocket of unmaintainable code. Across a large codebase, those pockets accumulate into a structural problem that is expensive to address.
This is not a theoretical risk. Projects that have adopted AI coding tools aggressively report higher rates of code churn, more time spent debugging AI-generated code, and difficulty onboarding new developers who cannot trace the reasoning behind existing implementations. The tools accelerate output. They do not necessarily accelerate understanding.
The charlatan problem
The ecosystem around AI agents includes a mix of genuine tooling and marketing-driven products that overpromise and underdeliver. The subsidized model access, where companies sell compute at a loss to acquire users, creates a temporary illusion of cost-effectiveness that will not survive when prices normalize. The agent frameworks that orchestrate multiple model calls add complexity and latency that may not be justified by the improvement in output quality. The vendors who claim their tools "replace developers" are selling to decision-makers who do not understand what the tools actually do.
The post's frustration is that this ecosystem makes it harder to have a clear conversation about what AI tools are actually good for. The narrow, genuine benefits get drowned out by grandiose claims about autonomous agents and the end of traditional programming. Developers who want to use these tools effectively have to wade through layers of hype to find the parts that are worth their time.
What remains after the hype
The useful nuggets are real, even if they are smaller than the marketing suggests. AI models are good at pattern matching, code explanation, and generating boilerplate. They are bad at architectural decisions, understanding business requirements, and reasoning about systems they have not seen before. The most productive workflows treat the model as a tool that accelerates specific tasks, not as a replacement for the developer's judgment about what to build and how to build it.
For developers wondering what is worth paying attention to, the answer is simple. If a tool helps you write code you understand, faster than you could write it alone, it is useful. If a tool generates code you do not understand, or makes architectural decisions you cannot verify, it is a liability. The distinction is not about the tool. It is about whether the developer remains in control of the output.