AI Architecture Updates: September 1, 2026

1. Moving Coding Agents Off the Laptop and Into a Governed Cloud Runtime

InfoQ / DoorDash. Writing on InfoQ, Leela Kumili reports on Flux, DoorDash’s platform for running engineering agent workloads in the cloud rather than on developer laptops, which automated roughly 130,000 tasks in a single month and now handles more than 25,000 automated code reviews per week across 300-plus playbooks. The architecture rests on four primitives: Firecracker micro-VM sandboxes with sub-five-second 95th-percentile setup, an in-house MCP Agent Gateway that enforces scoped permissions, audit logging, quotas, and a kill switch, YAML playbooks that mix agent-driven steps with deterministic code and validation rules, and multiple invocation surfaces spanning Slack, GitHub, cron, and CLI. The pattern matters because it reframes agent reliability as an infrastructure problem, treating isolation, identity, and observability as first-class runtime concerns instead of properties bolted onto an assistant. Source