Vercel Sandbox Drives: terabyte persistence, snapshots for agents
Your agent can finally keep a workspace between disposable runs and clone it across a fleet. Also: Copilot CLI auto-routes models and drops MAI-Code-1-Flash.
Copy markdown
Drives get snapshots: one workspace, many sandboxes
@vercel/sandbox 3.3.0 adds terabyte-scale persistent drives you mount into otherwise-disposable sandboxes, plus point-in-time snapshots. A read-write drive attaches to one sandbox at a time; a read-only snapshot mounts across many sandboxes at once. Data outlives the sandbox that made it.
Build this weekend: an agent that remembers
Clone your repo and install dependencies once onto a drive, snapshot it, and boot every future run from that warm state instead of rebuilding. Pair it with the new managed 'universal' image (Node 24, Python 3.14) and multi-region failover to fan a long-horizon coding agent out from a single snapshot.
Drive management lands in the CLI
The sandbox CLI (4.4.0) adds `sandbox drives get-or-create` and mount commands, so you can wire persistent storage into a script or CI job without touching the SDK. It also ships opt-out telemetry that flags AI-agent traffic.
Elsewhere: Copilot CLI auto-routes models, deprecates MAI-Code-1-Flash
GitHub's Copilot CLI now does adaptive model orchestration ('Project HydraFusion'), choosing a model per step, and gained a Jira integration in the Copilot app. In the same wave GitHub deprecated MAI-Code-1-Flash across every Copilot surface — if you pinned it, repin now.