Aishwarya Srinivasan

EP 42 Architecting AI Agents: The Shift from Models to Systems | Aishwarya Srinivasan

• One of the biggest prototype-to-production gaps is treating LLMs as tools versus building tools with LLMs inside them. As Aishwarya Srinivasan puts it, the “magic box” wrapper era of 2022–2023 is over — production agents are “way more of a software engineering challenge rather than just a model building challenge,” orchestrating small, mid, and large models, memory, and tool calls as one system. Episode takeaways
• “How much autonomy is right autonomy” has no universal answer — it depends entirely on the user journey, the points of risk, and the level of compliance you’re building under. The job is identifying exactly where a human-in-the-loop or a deterministic checkpoint belongs, not picking a global setting. Episode takeaways
• Quantify every step. Srinivasan’s core reliability principle is to make each step of the system measurable and logged — for a coding copilot, that means separate quantification for writing code, reviewing it, writing unit tests, and merging to main. Without checkpoints and traces, when something breaks you can’t reverse-engineer what broke, why, or how to stop it recurring. Episode takeaways