AI Dev 26 x SF | Tushar Jain: Shipping Agents Safely, Boundaries That Actually Work

Agents can write code, call APIs, install packages, and modify files. If you've built with them, you've already encountered the core challenge: they take real actions on real systems, and the surface area for failure grows with every capability you give them. Prompt guardrails shape intent but they don't enforce it. Once an agent has write access to production infrastructure, you need runtime controls, not just good instructions. In this talk, Tushar Jain from Docker draws on production experience to walk through practical patterns for scoping what an agent can touch and for how long, isolating execution so a failure stays contained, managing credentials without over-provisioning access, and building recovery into the loop rather than treating it as an afterthought. You'll leave with a concrete mental model for agent permissions and isolation, and patterns you can apply regardless of which framework or model you're building on.