we reimplemented lvm for our microvm hypervisors, with worse guarantees on purpose. its ~100ms volume-group lock was fine weekly but brutal at 200 launches/sec. keeping device-mapper, new control plane, ~100x faster allocation.
- noisy test failures burn hours: people reread the same logs, retry jobs to see if a failure disappears, and slowly stop trusting failures that point at something worth fixing. one approach: give each failure a stable identity you can compare over time.
- an agent can hand you a fix that works and passes its tests but could still be the wrong call for your system. you only catch that if you've kept enough context to challenge it. iris on where she stays involved to keep that context.
- when the agent's code breaks at 2am, you're the one on call, not the agent. here's a practical ai-assisted workflow for staying responsible for what ships: linter guardrails, reference tests, and reviewable prs.
- cloud-native worked because we built on a few composable primitives: compute, storage, networking. kyle's take: software delivery needs the same shift. he names the six infrastructure primitives to build on.
