Architecture AI Updates: November 21, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. Martin Fowler’s Fragments Cite CodeScene Data on Review Token Costs and Steve Yegge’s Case for Human-Speed Agents
Martin Fowler published a November 20 Fragments post that includes a CodeScene experiment reported by Marcus Borg on whether healthy code is cheaper for agents to review. Across three LLMs, 465 pull requests and 246 open source repositories run through a five-agent review and fix pipeline, a typical unhealthy pull request (CodeHealth 2.8) cost 22% to 35% more in tokens than a healthy one (9.7), with Complex Method and Complex Conditional smells as the main drivers. The post also relays Steve Yegge’s argument that “software factories” should be paced to human speeds, which Fowler supports by saying agents should run at a speed humans can govern. It also cites a Thoughtworks report line that code generation is no longer the bottleneck and verification is. Source
2. Google Donates Agent Substrate, a Kubernetes Layer That Separates Agent State From Compute, and gVisor to the CNCF
InfoQ reported that Google is donating Agent Substrate and gVisor to the Cloud Native Computing Foundation for vendor-neutral governance. Agent Substrate models each agent session as an actor that holds filesystem, memory and execution state, and assigns it to a reusable pre-warmed worker from shared Kubernetes pools when it resumes, then saves state and releases the compute when it goes idle. Google estimates agents spend 90% to 95% of their time idle while holding state, and says they can resume in under 500 milliseconds. gVisor adds isolation with a user-space application kernel that intercepts system calls, and Agent Substrate can use Cloud Hypervisor microVMs for workloads that need hardware-level virtualization. Source