Fixing a small issue like "search summary can't serve as body content" requires editing across Prompt, Tool, Runtime, Evidence, and Recovery — all the way through the stack. This is a field report from an Agent (a program that lets AI autonomously complete multi-step tasks) engineer on Juejin. We've noticed: once these AI applications pass the demo stage, what really blocks teams isn't how strong the model is — it's "how many layers do you have to touch to fix one bug."
What this is
The author lays out the common five layers: Prompt, Tool, Runtime, Evidence, Recovery. Each layer has seemingly clear responsibilities, but when you sit down to write code you discover that a single fact — for example, "was this web page actually read" — can be redefined across all five layers. Fixing one bug naturally means moving horizontally.
His core judgment: architectural maturity isn't about layers multiplying — it's about the blast radius shrinking when the same kind of problem hits. The specific approach is to reclaim the "right to define facts" to a single layer, so other layers can only consume its projection. The Tool Host alone proves whether a call actually happened; Runtime alone proves whether a permission is still valid; the model can only judge "is the evidence sufficient."
He provides measured numbers: general research request serialization size dropped ~31%, fund scenario ~12%, portfolio scenario ~24%.
Industry view
Supporters say: this is the watershed for whether Agent products can make it to production. In 2024 everyone was busy shipping demos; in the second half of 2025 the competition is maintainability — the faster and cheaper you can fix bugs, the more likely iteration can take off.
But there are dissenting views. First, models themselves are getting stronger — this generation may make low-level problems like "summary as body" disappear on their own, so architecture may not be the main contradiction. Second, "strict layering + single definition authority" demands high engineering capability from teams, which actually reinforces the moat for established large companies; small teams may not be able to afford it. Third, the author himself admits that the numbers above only prove requests got smaller — they don't directly prove improvements in tokens, latency, or answer quality. Clean architecture doesn't equal business effectiveness.
Impact on regular people
For enterprise IT: if you're selecting an AI vendor, you used to evaluate demos; next you need to evaluate "how quickly can it be fixed after something breaks, and does it require a rebuild from scratch." Maintenance capability will become the new procurement criterion.
For individual careers: "Knowing how to use AI" and "being able to maintain AI systems" are two different skills. The latter is rarer and more valuable. People who understand engineering and can make architectural judgments are becoming the new must-have.
For the consumer market: regular users won't notice a difference in the short term, but within a year the divergence in product stability and answer accuracy will be obvious. Products willing to invest at the foundational layer will pull away from competitors that "look about the same."