[KongchangAI]
· 2 min read· 1,411 words

Is AI Agent Output Being 'Enshittified'? How Much Time Do We Have Left?

Is AI Agent Output Being 'Enshittified'? How Much Time Do We Have Left?

AI Agent output may repeat the platform "enshittification" trajectory; developers need to establish quality standards and observability in advance.

Starting from a Hacker News question with only 5 upvotes, the article borrows Cory Doctorow's concept of "enshittification" to explore whether AI Agent output is heading toward quality decline. The author identifies three main paths: commercial incentives quietly penetrate through ranking weights and prompt design; over-alignment reduces output to "correct nonsense"; and multi-Agent chaining dilutes quality through cascading errors. The article emphasizes that Agents iterate far faster than search engines and social platforms, so the quality inflection point may arrive earlier and more covertly. Developers should actively counter this trend before Agents become the default entry point by recording chain observability, incorporating "usefulness" into evaluation criteria, and prioritizing transparent infrastructure.

Starting from a Hacker News Post

A short but pointed question appeared on Hacker News: "How much time do we have before Agent output gets enshittified?" The post received only 5 upvotes and 2 comments, but it touched on a widespread unease in today's AI community: is the content generated by AI Agents repeating the same fate as search engines, social platforms, and e-commerce platforms—gradually sliding from being useful to users toward being favorable to platforms and dismissive of users?

The term "enshittification" was coined by writer Cory Doctorow to describe the cycle in which platforms first court users, then court business customers, and ultimately exploit both ends until quality collapses. Applying this framework to AI Agents makes the problem very concrete: when Agent output increasingly depends on model vendors' commercial goals, ad injection, affiliate links, or over-filtering through "safety alignment," what users receive may no longer be honest answers but carefully curated results.

Cory Doctorow and the Original Context of "Enshittification"

"Enshittification" is a concept systematically articulated by Cory Doctorow in a 2023 article published in Wired, originally used to describe the decay trajectory of two-sided platforms such as TikTok, Facebook, and Amazon. Its core mechanism is: a platform first attracts users with a high-quality experience, and after achieving scale, shifts toward favoring business customers (e.g., increasing ad placements, adjusting algorithmic weights), and finally extracts value from both users and business customers simultaneously until the experience breaks down and users leave. Doctorow emphasizes that this is not a moral issue of individual companies but an inevitable incentive outcome under monopolistic structures—once a platform locks in users, it loses the competitive pressure to maintain quality. The reason this framework can be transplanted to AI Agents is that Agents likewise play the role of a "two-sided intermediary": one end connects to user needs, and the other end connects to model vendors, content providers, and commercial partners. When user switching costs are high and substitutes are few, Agent vendors have the same incentive as platforms to sacrifice output neutrality in exchange for commercial returns.

Three Paths to Agent Output Enshittification

The Penetration of Commercial Incentives

The biggest similarity between AI Agents and search engines is that they are both "middlemen of answers." Once an Agent becomes the default entry point for users to obtain information, the platform has a strong incentive to tie recommendation results to commercial interests. For example, a shopping Agent may prioritize products that pay commissions, and a travel planning Agent may insert partner hotels by default. This tendency does not appear in the explicit form of "advertising" but is hidden in ranking weights and prompt design, making it hard for users to detect.

"Correct Nonsense" Caused by Over-Alignment

Another path comes from safety and compliance. To make models "not say the wrong thing," vendors add a large number of refusal rules and disclaimers. As a result, the content given by Agents becomes increasingly bland and safe: no clear stance, no specific data, no actionable advice—only correct but useless boilerplate. This kind of "safety" is itself a form of enshittification—it protects the platform while sacrificing users' actual benefits.

The Specific Meaning of "Alignment" in Agent Scenarios

Alignment originally refers to making a model's behavior conform to human intentions and values. Common methods include RLHF (reinforcement learning from human feedback), constitutional AI, and refusal rules in system prompts. At the chatbot stage, alignment mainly affects "whether to answer"; but in Agent scenarios, alignment has a deeper impact—it determines which tools the Agent calls, which sources it retrieves, and how it organizes multi-step reasoning. For compliance and reputation reasons, vendors tend to set conservative default policies, such as attaching disclaimers to all questions involving medical, financial, or legal matters or directly referring users elsewhere. The problem is that the benefits of this conservative strategy accrue to the platform (reducing risk), while the costs are borne by users (receiving useless answers). More subtly, over-alignment may also cause Agents to proactively abandon valuable exploration paths in multi-step tasks, choosing only "safe but mediocre" solutions, thereby systematically lowering output quality.

Quality Dilution in Multi-Agent Collaboration

When Agents begin to call each other and chain together to complete tasks, each layer may introduce noise and bias. If a retrieval Agent returns low-quality sources, the downstream summarization Agent will amplify this noise into seemingly authoritative conclusions. The longer the chain, the lower the fidelity of the original information, and the more the final output resembles a game of telephone.

Multi-Agent Architecture and Error Accumulation

Multi-Agent systems usually consist of roles such as planner, retriever, executor, and summarizer, connected through message passing to complete complex tasks. The advantage of this architecture is clear division of labor and parallelism, but it also introduces the problem of "error accumulation": each layer of Agent has its own prompt bias, tool limitations, and hallucination risks, and errors are gradually amplified along the chain rather than canceling each other out. Academia often uses "cascading errors" to describe this phenomenon—an upstream retriever returns a low-quality source, a midstream summarizer packages it into a seemingly authoritative claim, and a downstream generator then produces specific recommendations based on it, so the final user sees a logically coherent but factually wrong conclusion. The longer the chain, the lower the fidelity of the original evidence. Mitigation methods include: setting up cross-validation Agents at key nodes, requiring each step's output to include source citations, and scoring intermediate results for confidence. Only by understanding this mechanism can one judge whether the quality decline comes from insufficient capability of a single model or from systemic flaws in the architecture itself.

Why Discuss This Timing Now

The "how much time" in the post is not groundless worry. Current Agent products are at a critical point in the transition from demos to scale: on one hand, tool calling, long context, and memory capabilities enable Agents to handle increasingly complex tasks; on the other hand, commercialization pressure is rising in tandem. Historically, it took search engines about ten years to go from "useful" to "dominated by advertising," while the quality decline of social feeds was even faster. Agents iterate far faster than either of these, which means the quality inflection point may arrive earlier and more covertly.

Interestingly, although there were only two comments, this kind of discussion recurs in developer communities, indicating that the concern is no longer an isolated case. What developers really care about is not whether Agents will make mistakes, but whether they are still "on the user's side."

What Developers Can Do

Individual developers are not powerless against Agent enshittification.

Maintain observability of output. Record the input and output of each stage in the Agent chain, especially retrieval sources and ranking rationale. When output quality declines, you can quickly identify whether it is a model problem, a data problem, or a product strategy problem.

Write "usefulness" into evaluation criteria. In addition to accuracy, evaluate the specificity and actionability of output. An Agent that only says "recommend consulting a professional" may be very safe on metrics, but it is worthless to users.

Prioritize transparent infrastructure. Although open-source models and self-hostable Agent frameworks have higher maintenance costs, they can prevent vendors from quietly inserting commercial preferences into prompts and ranking layers.

Conclusion

Will Agent output be enshittified? The answer is likely "yes"; the only question is speed and degree. The value of this Hacker News post with only 5 upvotes is not in providing a conclusion but in reminding us: before Agents become the default entry point, establish quality standards and observability first. Waiting until output has already gone bad to look back will be far more costly.

One-line Summary

AI Agent output may repeat the platform "enshittification" trajectory; developers need to establish quality standards and observability in advance.

Paragraph Summary

Starting from a Hacker News question with only 5 upvotes, the article borrows Cory Doctorow's concept of "enshittification" to explore whether AI Agent output is heading toward quality decline. The author identifies three main paths: commercial incentives quietly penetrate through ranking weights and prompt design; over-alignment reduces output to "correct nonsense"; and multi-Agent chaining dilutes quality through cascading errors. The article emphasizes that Agents iterate far faster than search engines and social platforms, so the quality inflection point may arrive earlier and more covertly. Developers should actively counter this trend before Agents become the default entry point by recording chain observability, incorporating "usefulness" into evaluation criteria, and prioritizing transparent infrastructure.

#AI Agent#enshittification#output quality#multi-Agent#alignment#observability
Share:

Related articles