Vercel Sandbox observability now reports memory usage in the dashboard and through the vercel metrics CLI command. The dashboard’s Memory Usage card shows average, P75, and P95 values across sandboxes, and detail charts scale to a sandbox’s memory limit with a dashed reference line at 85%. The source also says that memoryUsedBytes is available in the Observability query builder and can be used to create alerts.
The mechanism is that memory telemetry is exposed at three layers: a fleet-level percentile view, a limit-aware view for an individual sandbox, and a queryable CLI and observability measure. This makes it possible to distinguish a sustained rise in memory consumption from occasional peaks, then route the resulting condition into a custom query or alert instead of treating memory exhaustion solely as a post-failure debugging problem.
Why now: sandboxed workloads, including automated and agent-driven tasks, can vary substantially in duration and resource profile, so operators need a way to see approach-to-limit before a hard ceiling becomes the first visible signal. The first implication is that teams can set policies around memory headroom using actual workload distributions rather than only configured limits. The evidence boundary is that Vercel describes the new measurements and alerting path, but does not provide evidence here about alert latency, retention, pricing, or the rate at which these signals reduce failures. The first watch condition is whether teams can use this metric to establish reliable pre-exhaustion alerts across bursty sandbox workloads.