Kubernetes Is Not Your Sandbox#
AI Engineer World's Fair 2026 digest Program Core pick 29 June – 2 July 2026
Speakers: Ivan Burazin
- Why: Teams are reaching for Kubernetes to run agent sandboxes, and the mismatch — an orchestrator built to hold steady state versus environments that are born, forked, and killed in under a second — has a measurable, benchmarked cost. Directly relevant to K8s architecture and operations.
- What: A benchmark running the same 50-action bug-fix trajectory across five stacks, yielding a 12x spread — 12.9s on the fastest, 161.5s on the slowest — with the gap attributed to per-action lifecycle overhead, not compute. Every stack is named and each number's mechanism explained. The talk also frames the requirements contradiction: sandboxes went from fast code-snippet runners to stateful long-running agent boxes to ten thousand ephemeral environments that fork for RL rollouts and die in under a second — all at once (from abstract only).
- How: Cross-stack execution-time decomposition isolating lifecycle overhead (scheduling, environment start/teardown) per action from actual compute.
- Where applicable: Any team running agent or RL-rollout environments at scale; the comparison is stack-agnostic and directly informs substrate choice. The abstract closes with "wdyt?", suggesting the talk invites debate — worth attending for the numbers and mechanisms regardless.