Vercel shipped four small AI Gateway updates on July 31. Read together, they are more useful than any one changelog line: AI Gateway is turning into an operating surface for agent inference.
The budget release is the clearest control. Spend budgets can now attach to a team or project, not only to an API key. A request can fall under multiple budgets, and it has to pass all of them. If any applicable cap is exhausted, the gateway rejects the request until the budget resets or changes. Alerts at 50%, 75%, and 100% are informational; the budget limit is the blocking mechanism.
Fast mode is the latency side of the same idea. Instead of hardcoding provider-specific fast variants everywhere, callers can set speed: "fast" under gateway options. The gateway serves the fast tier when routable and falls back to standard speed when it is not. Direct fast slugs still exist for configurations that need explicit fallback lists.
The Laguna S 2.1 capacity increase matters because agent workloads are spiky and long-running. Vercel says both poolside/laguna-s-2.1 and poolside/laguna-s-2.1-free now have 10x more Gateway capacity, aimed at higher-volume agentic coding and longer tasks.
The MCP update closes the loop. Vercel MCP now supports the 2026-07-28 specification, giving newer clients a stateless request model and updated authorization behavior while keeping 2025 clients on the same endpoint.
For New Runtime, the pattern is consistent with earlier Vercel posts: the agent platform is becoming a bundle of boring controls. Spend, latency, capacity, and protocol compatibility are not side panels. They decide whether an agent loop can run unattended without turning into a surprise bill, a queue, or a broken tool connection.