korrents

korrents · video

2,000+ Hours of Learning Agentic AI System Design in 20 Minutes

Aishwarya Naresh Reganti · 20m · youtube.com ↗

15 korrents from this recording

Aishwarya Naresh Reganti did not write this page.

Every claim below is a statement made in this recording, quoted word for word and linked to the second it was said, so you can hear it rather than take our word for it. The wording comes from the transcript published alongside the recording; the sentence above each quote is our reading of the claim, not their wording.

  1. 0:06 · watch on youtube.com↗

    And the candidates who don't get hired often have the same thing in common. Yes, they know what rag is, MCPs, or even tool calling is, but struggle to apply those concepts in a real business problem, or explain what it takes to scale an AI system to production.
  2. 1 min later
  3. 1:33 · watch on youtube.com↗

    remember that you need to understand the scope of the problem very well before going into solutioning. A lot of people get confused in this, but a huge part of system design is simply scoping the problem and making your assumptions explicit.
  4. 2 min later
  5. 3:20 · watch on youtube.com↗

    And if you're answering a system design interview question, this is where you want to ask your interviewer as many questions as possible, because the more assumptions you make up front, the better your system design or your architecture design will get.
  6. 3 min later
  7. 6:27 · watch on youtube.com↗

    If our evals later show that many questions actually need information across several documents, then we can revisit that decision, but do not overcomplicate your solution.
  8. 1 min later
  9. 7:22 · watch on youtube.com↗

    So, for one, permissions, pretty much follow least privilege. Every tool should get only the access to the information that it requires.
  10. 7:46 · watch on youtube.com↗

    So, higher consequence actions like refunds or cancellations should sit behind a deterministic policy and should require human approvals.
  11. 1 min later
  12. 8:17 · watch on youtube.com↗

    Now, user messages, retrieve documents, and tool outputs should all be treated as untrusted data. For instance, just because your system sees an instruction inside a document, it shouldn't just go and execute it, especially once the agent has access to private data and tools that can take actions.
  13. 1 min later
  14. 9:34 · watch on youtube.com↗

    Now, remember that just because these four exist doesn't mean you have to use all of them.
  15. 2 min later
  16. 11:16 · watch on youtube.com↗

    If something goes wrong in production, you shouldn't just know the final answer was bad, you should be able to trace where it went wrong and have enough information to actually fix it in the next go.
  17. 1 min later
  18. 12:01 · watch on youtube.com↗

    And the most important thing about evals is they need to be actionable.
  19. 2 min later
  20. 13:45 · watch on youtube.com↗

    something more subjective, like whether an answer is grounded in its context, whether it's useful, whether it's matching the brand voice of the company, those are the cases where you need AI judges or LLM judges.
  21. 3 min later
  22. 16:50 · watch on youtube.com↗

    because that's where interviews are really looking for signal as to how well you can understand different design decisions and make architecture tradeoffs. Now, that layer is what separates your great candidates from good candidates.
  23. 2 min later
  24. 18:45 · watch on youtube.com↗

    Moving to a stronger model can be one of the options, but that's only if the model is your bottleneck.
  25. 19:01 · watch on youtube.com↗

    Now, one agent might have too much context and too many tools, in which case having specialist agents behind an orchestrator might actually earn its complexity.
  26. 19:10 · watch on youtube.com↗

    And this is pretty much what interviewers are looking for, not someone who says at the first shot that they'll use a multi-agent system. Every complexity has to be earned, and you need to explain why you made those decisions.