AI SDKs converge on agent tools while platforms harden builds and sandboxes

Agent tools reshape AI SDK search patterns

The week's dominant story is a coordinated shift from search parameters to agent tools across Vercel's AI SDK providers. Vercel's xAI provider deprecated searchParameters in favor of web_search and x_search agent tools, adding an enableImageSearch option to xai.tools.webSearch() that passes through to the xAI Responses API as enable_image_search. The stable xAI v3.0.93 channel received the same treatment simultaneously. Meanwhile, Azure's OpenAI provider gained a web search tool for the Azure OpenAI Responses API, mirrored in the stable Azure channel. The Google Vertex provider added support for gemini-embedding-2 and two Deep Research preview models (deep-research-max-preview-04-2026, deep-research-preview-04-2026). To round out the model additions, AI Gateway support landed for Qwen 3.7 Plus from Alibaba and MiniMax M3 with its 1M-token context window.

On the MCP front, both the @2.0.0-canary.56 and @1.0.45 channels fixed MCP ping responses to return an empty result per the JSON-RPC spec, closing a long-standing spec compliance gap. The AI SDK core also introduced standalone stream transformation helpers (toUIMessageChunk, toUIMessageStream, toTextStream) that decouple stream conversion from streamText result objects, deprecating the result-bound methods for removal in the next major release.

Build resilience and infrastructure hardening

Vercel's elastic build machines now monitor memory usage and automatically upgrade machine tiers to prevent out-of-memory failures — a significant reliability improvement for teams hitting OOM errors during deployment. The system also avoids downgrading fast but memory-intensive builds. On the pricing side, function invocations moved to per-unit billing ($0.0000006 per invocation) for Pro customers, smoothing cost scaling across usage patterns.

The Vercel CLI v54.6.1 improved the vercel env add flow by asking about sensitivity before collecting values, and added experimentalServices discovery from Build Output API config during vercel build. The Python builder gained bytecode compilation support for hive deployment paths, while @vercel/express bumped @vercel/nft to 1.10.0 with moduleSyncCatchall tracing enabled.

Sandbox, storage, and developer experience leaps

Cloudflare Agents sandboxes can now expose services through Cloudflare Tunnel via sandbox.tunnels.get(port), creating zero-config *.trycloudflare.com preview URLs without requiring a Cloudflare account or custom domain. Cloudflare Realtime's WebSocket adapter gained automatic reconnection and WebRTC media buffering, making it production-ready for live audio/video routing across the global network. On the data side, D1 migrations now support nested layouts through migrations_pattern, aligning with Drizzle's folder structure. Workers added TypeScript types for AgentMemoryNamespace and composite feature types, signaling continued investment in the Workers agent model.

Vercel Sandbox now supports Docker containers inside sandboxes, enabling containerized test dependencies, image validation, and CI workflows without host-side Docker. Railway's CLI v4.66.0 introduced railway sandbox commands (create/list/ssh/exec/destroy), and Railpack v0.24.0 now respects package manager versions from mise configuration files.

Cloudflare Security Center increased scan frequency — daily for Enterprise, every 3 days for Pro/Business, weekly for Free — with on-demand scans now available on any plan. Logpush datasets gained Turnstile Events and field additions for DEX Device State, Gateway HTTP, and HTTP requests.

Across the collection, Next.js canary builds progressed with Turbopack analyzer improvements, image optimization operation cache, and 'use cache' bail-out fixes, while Turborepo added auth HTTP timeouts and fixed Windows graceful shutdown.