Vercel AI Updates: May 12, 2026

1. Sandbox firewall gains request proxying and filtering for agent egress

Vercel. Vercel Sandbox can now route outbound traffic through a customer-supplied proxy with matchers on path, method, query, and headers, and forwarded requests carry a vercel-sandbox-oidc-token identifying the team, project, and sandbox. For teams running coding agents inside Sandbox, this is the missing piece for egress logging, redaction, and policy enforcement without rewriting the agent itself. The feature is in beta for Pro and Enterprise. Source

2. AI SDK workflow adds flexible tool descriptions

Vercel. @ai-sdk/workflow@1.0.0-canary.47 introduces a “flexible tool descriptions” feature alongside dependency bumps to ai@7.0.0-canary.131 and @ai-sdk/provider-utils@5.0.0-canary.38. The change lets workflow authors vary how tools advertise themselves to the model per call site, which matters when the same tool needs different framing inside different sub-agents. Source

3. AI SDK workflow exposes explicit step numbering in streaming

Vercel. Canary 48 follows up by adding a stepNumber option to doStreamStep, so callers can produce StepResult objects with the correct index instead of relying on the runtime’s implicit counter. It is a small API surface change but a meaningful one for anyone wiring AI SDK workflows into external orchestrators or persisting step state across runs. Source

4. Bedrock provider stops forwarding unsigned reasoning content

Vercel. @ai-sdk/amazon-bedrock@5.0.0-canary.58 ships a one-line fix to skip passing unsigned reasoning content back to Bedrock. The same fix backports to the stable 4.0.104. If you have seen Claude-on-Bedrock requests fail after a tool call when extended thinking is on, this is the patch that quietly unbreaks them. Source

5. xAI provider tracks the new openai-compatible canary

Vercel. @ai-sdk/xai@4.0.0-canary.58 is a dependency-only bump to @ai-sdk/openai-compatible@3.0.0-canary.45 and @ai-sdk/provider-utils@5.0.0-canary.38. No new feature surface, but it keeps the xAI provider current with the shared OpenAI-compatible transport layer the SDK is consolidating on. Source