Run Terminal-Bench and other Harbor evals on Vercel Sandbox

In its September 28-October 2 roundup, Google Cloud highlights storage-management guidance, an Apigee X blueprint that routes AI requests by complexity, and a new agent-evaluation session. Google describes pairing standardized traces with model-based judges and deterministic checks to detect quality regressions. The roundup also points to storage operations using Storage Intelligence Advisor and Batch Operations, and advertises an Agentic Data Cloud event for November 4.

Google Cloud has made Spanner queues generally available, letting applications commit database changes and queued tasks together in one transaction. Messages support delayed delivery, SQL consumption and transactional acknowledgments for agent handoffs and other asynchronous workflows. Google specifies at-least-once delivery and at-most-once acknowledgment. External calls still need safeguards such as the idempotency key used in Google's example.

Google Cloud is previewing CPU startup boost for GKE, which temporarily increases container CPU requests and lowers them after readiness without restarting the container. It requires GKE 1.36.0-gke.4447000 or later; Standard clusters need Vertical Pod Autoscaling enabled, while Autopilot enables it by default. Boost policies support multipliers or fixed additions, and Autopilot workloads must still satisfy CPU-to-memory ratios.

Vercel says its latest AI SDK for Python adds an experimental evaluate() API for Jev, a model that returns structured classification decisions and confidence estimates. The accompanying tutorial explores distinguishing Python from English and constructing code through syntax-tree choices. Its examples still misclassify some inputs and produce mostly incorrect code, so the article urges developers to test whether Jev is accurate enough for their specific tasks.

Cloudflare has closed a network-performance incident affecting Madrid after reporting a fix on September 30. Its incident timeline records monitoring from 14:06 UTC that day and a resolved update at 05:42 UTC on October 2. The provider classified the incident as minor; its notice does not identify a root cause or quantify customer impact.

Cloudflare marked its Workers build-delay incident resolved at 21:27 UTC on October 1. The company had opened an investigation at 15:33 UTC on September 30, warning of potential effects on multiple customers, then moved to monitoring after implementing a fix at 17:42 UTC on September 30.