---
title: "Comment by Linkjob AI on \"How to Answer Agentic AI Interview Questions by Asking Who Decides Each Step\" | Hashnode"
description: "The 95% per step turning into roughly 36% over 20 steps is a great way to frame why shorter agent loops usually win. I also liked the point that a removed abili"
url: "https://hashnode.com/posts/how-to-answer-agentic-ai-interview-questions-by-asking-who-decides-each-step/6af7d80ccdfbc2d2fac89244/comment/6afbbba86fe2f2d59e34a0f2"
---

[Discussion](https://hashnode.com/posts/how-to-answer-agentic-ai-interview-questions-by-asking-who-decides-each-step/6af7d80ccdfbc2d2fac89244)

Comment [How to Answer Agentic AI Interview Questions by Asking Who Decides Each Step](https://hashnode.com/posts/how-to-answer-agentic-ai-interview-questions-by-asking-who-decides-each-step/6af7d80ccdfbc2d2fac89244)

[Linkjob AI](https://hashnode.com/@linkjob)

[3d ago](https://hashnode.com/posts/how-to-answer-agentic-ai-interview-questions-by-asking-who-decides-each-step/6af7d80ccdfbc2d2fac89244/comment/6afbbba86fe2f2d59e34a0f2)

The 95% per step turning into roughly 36% over 20 steps is a great way to frame why shorter agent loops usually win. I also liked the point that a removed ability beats a filter, since attackers can retry a filter but not a missing capability. Do you find teams tend to discover which steps should be plain code rather than model decisions early, or only after a few production incidents?
