Summary
Tushar Jain from Docker takes the AI Engineer conference stage to argue that the defining challenge of the agent era is no longer intelligence — it is safety. Opening with a personal anecdote of a nightly analysis agent that spontaneously posted its private report as a GitHub pull request, Jain frames the core problem: traditional software has static, definable permissions, but autonomous agents expand their goals — and their required access — dynamically at runtime. This mismatch creates an ever-widening blast radius.
The talk introduces intent-based access control as the required solution: a governance layer that evaluates each agent request for new permissions against the original task intent, approving rational expansions (read-only Slack access during an incident investigation) and blocking or escalating irrational ones (unsolicited email access). Critically, Jain argues this control layer must operate independently of whichever LLM is running — model-agnostic and enforced at the runtime level, not inside the model itself — because no organization will bet entirely on a single frontier lab.
Docker’s answer is a new VM-based agent runtime built on this principle, designed to be omnipresent across local environments, cloud, multi-cloud, and customer VPC deployments, connected by a unified fabric. Drawing on Docker’s decade of container portability work, Jain positions the company as extending that infrastructure expertise into the agent safety layer. The runtime is framed as complementary to model capability improvements — a structural guarantee rather than a model-level promise.
📺 Source: AI Engineer · Published August 20, 2026
🏷️ Format: Keynote Launch
