1 min lesson
The fifth class is the AI-native one
Connect "The fastest debugging question is 'which layer owns this'?" to the decision you would make.
Step 1 of 2
The fifth class is the AI-native onemodel regression
The class a non-AI QA background will miss entirely is the model regression. A model upgrade or a backend routing change silently shifts behavior, diff quality or latency while the app version is unchanged. Nothing crashes. Users just feel that "it got dumber," and your only tell is a behavior change correlated with a model swap rather than a release.
Sort every failure report into client, context/retrieval, agent orchestration, network or model. The fastest debugging question is "which layer owns this?" A hallucinated API is model/generation; an ignored file is context/retrieval; a streaming stall is network or backend. The layer tells you which logs to pull and which team to hand it to.
Learn more
Full explanation
Which Layer Owns the Bug
Interactive diagram. Tab through its regions; each focused region shows its detail in the panel below.
Name the layer before you investigate. It points you at the right logs and the right team.
When given an AI-product bug, classify before you investigate: "This reads like a context-retrieval failure, not generation, so I'd check indexing and the actual @-context payload before I blame the model." Naming the layer first shows you hold the architecture in your head: a VS Code fork, a client talking to model backends, retrieval feeding context.
Don't reflexively blame the model. Most loud "the AI is dumb" reports root-cause to context or retrieval: the model never saw the file, so it had nothing to act on. Blaming the model when the bug is in indexing sends the report to the wrong team and burns a day.
Learn more
Optional practice
Test yourself on The fifth class is the AI-native one
QA user says Cursor "ignored the file I had open and gave me a totally generic answer." Which failure class is most likely and what do you check first?
QUsers report that Agent diffs got noticeably worse since Tuesday, but every report shows the same Cursor build they ran last week. Which failure class is most likely, and what do you correlate against?