Think Outside the Box, Deploy Inside It: AI Agent Control via Infrastructure
- Speaker: Priya Ananthasankar
- Room: CC 236
- Time: Sun 11:20 am – 11:50 am
- Format: Lecture (30 Min + Q&A)
- Difficulty: Introductory / Some experience required
- Track: Development & Dev Tools
- Additional Tags: DevOps, AI / ML, Creative
- Presenter Location: In-person
- Experience: several-th time speaking
Description:
What’s the balance between giving AI agents enough autonomy to be useful and keeping them under control? This session explores that tension through a practical experiment: deploying the same AI agent across progressively more constrained environments and observing how its behavior changes.
Priya walks through three stages:
- Unrestricted deployment — watching how an unconstrained agent explores and behaves
- Tightened constraints — limiting environment, permissions, or resources and analyzing the impact
- Minimal guardrails — identifying the smallest set of controls needed to ensure safety, reproducibility, and reliability without killing creativity
Through concrete examples, attendees will discover:
- How much freedom an agent actually needs to stay useful
- How far you can lock things down before losing value
- Where the real control points live — and why control is about *channeling* creativity, not caging it
This talk blends infrastructure, DevOps, and AI agent design to show how deployment environments shape agent behavior as much as model architecture.
Target Audience:
- Cloud engineers
- DevOps engineers
- Infrastructure engineers
- AI practitioners