Vercel and Cloudflare converge on agent-ready infrastructure

Summary

The 2026-07-31 signal set was dominated by two platform players moving in parallel: Vercel (6 signals) and Cloudflare (4). Vercel broadened AI Gateway into a provider-neutral, multimodal model catalog and hardened enterprise identity, while Cloudflare pushed browser-agent ergonomics and edge protocol/media tooling. The through-line across both is infrastructure maturing for production agents and multi-agent systems — first-class human handoff, per-agent isolation, keyless CI, and deployment-level SSO — rather than one-off model or feature launches.

Key themes

• AI Gateway as a provider-neutral multimodal catalog: Vercel added Thinking Machines' open-weights, 1M-context Inkling Small (MoE with adjustable thinking effort) and MiniMax H3 for 2K text-to-video, then cut GPT-5.6 pricing (~80% on Luna, ~20% on Terra) and sped Sol fast mode ~2.5x — all passed through with no token markup, positioning the gateway as neutral plumbing across modalities.
• Agent execution infrastructure maturing for production: Cloudflare Browser Run added a structured human-in-the-loop pause-and-resume handoff (a human steps in via Live View for MFA/CAPTCHA/credential steps, then the agent resumes), and Vercel Sandbox added per-agent Linux user isolation so multiple agents share one sandbox with private home directories — both shifting the design target toward reliable, orchestrated agent fleets.
• Enterprise identity and keyless access tighten: Vercel Passport reached GA, gating every deployment behind Okta, Microsoft Entra ID, Auth0, or any OIDC IdP ($100/project/month, no external-user cap), while Turborepo and Vercel Remote Cache adopted OIDC to replace long-lived CI cache tokens with short-lived workload identity.
• Cloudflare extends edge media and protocol tooling: a provisioning API for isolated Media over QUIC relays (permissioned publisher/viewer control across 330+ cities) targets sub-second live media as a WebRTC/HLS alternative, alongside response-phase Cache Response Rules that reclaim cacheability without origin changes and the open-sourced pvcli client for testing privacy protocols like OHTTP.

Notable items

• Vercel Passport reaches general availability — deployment-level SSO enforcement behind a team's own OIDC identity provider (Okta, Entra ID, Auth0), priced at $100 per project per month with no cap on external users.
• Cloudflare Browser Run ships a first-class human-in-the-loop handoff — agents formally pause and signal for help, a person completes the step via Live View, and the agent resumes, making long-running browser automation production-viable while keeping credential entry with a human.
• Vercel Sandbox adds multi-user isolation (createUser/createGroup/asUser) so multiple AI agents run as separate Linux users inside one sandbox, cutting cold-start and resource overhead for multi-agent systems while preserving per-agent filesystem isolation (JS SDK only).
• Vercel AI Gateway cuts GPT-5.6 Luna pricing ~80% and speeds Sol fast mode ~2.5x with no token markup — materially changing the cost math for high-volume agent loops routed through the gateway.
• Cloudflare launches a provisioning API for isolated Media over QUIC (MoQ) relays with per-relay publisher/viewer permissions, turning its edge MoQ fabric into programmable, standards-based low-latency streaming infrastructure.
• Turborepo and Vercel Remote Cache add OpenID Connect support, letting CI/CD pipelines authenticate to the Remote Cache with short-lived OIDC tokens instead of long-lived cache secrets — removing a common secret-sprawl risk for monorepo builds.

Source coverage

Source rows used: 10