---
title: "Comment by AI Developer on \"Here Be Monsters: What Do You Still Need to Know When Agents Do the Work?\" | Hashnode"
description: "The SonarQube anecdote shows well that the useful skill is detecting that something is wrong, not writing the fix. One thing I'd add to the \"just-in-time depth\""
url: "https://hashnode.com/posts/here-be-monsters-what-do-you-still-need-to-know-when-agents-do-the-work/6ae03c11e1c6135b06ad49ec/comment/6ae1e5b814c359e517385d96"
---

[Discussion](https://hashnode.com/posts/here-be-monsters-what-do-you-still-need-to-know-when-agents-do-the-work/6ae03c11e1c6135b06ad49ec)

Comment [Here Be Monsters: What Do You Still Need to Know When Agents Do the Work?](https://hashnode.com/posts/here-be-monsters-what-do-you-still-need-to-know-when-agents-do-the-work/6ae03c11e1c6135b06ad49ec)

[AI Developer](https://hashnode.com/@lycore)

[Oct 28](https://hashnode.com/posts/here-be-monsters-what-do-you-still-need-to-know-when-agents-do-the-work/6ae03c11e1c6135b06ad49ec/comment/6ae1e5b814c359e517385d96)

The SonarQube anecdote shows well that the useful skill is detecting that something is wrong, not writing the fix. One thing I'd add to the "just-in-time depth" idea is making those detectors explicit and independent of the agent: complexity budgets, dependency diff alerts, mutation testing on critical modules, and a short architecture-decision file the agent must update when it changes a boundary. They act as the alarms that tell you when to dive. The training question gets harder, though. Do you think juniors should deliberately debug agent-written code in a sandbox with seeded faults, so they build the pattern recognition that used to come from writing it by hand?
