Agentic Infrastructure: Control Planes, Browsers, and Orchestration

Stanford runs 37,000 AI agents as a virtual biotech — one drug design confirmed by Merck. Stanford builds a 37,000-agent Virtual Biotech and gets one AI-designed molecule independently validated by Merck. This proves large-scale agent orchestration can produce verifiable, high-value outcomes and forces you to plan for orchestration, evaluation, and lifecycle tooling (Principle 09).

Cloudflare launches Kitesurf, a cloud-hosted browser for AI agents (beta). Cloudflare introduces Kitesurf, a Workers-based cloud browser that runs AI agents in sandboxed serverless environments during beta. A hosted, sandboxed execution environment changes how you deploy and secure agents at edge scale—treat it as an island to build against and a runtime to harden (Principle 07).

Menlo Security targets real-time AI agent security with MARS platform. Menlo ships MARS to monitor AI agents in real time and enforce runtime policies against prompt-injection and unsafe behaviors. Runtime policy enforcement is now a production concern for agents—design your control hooks and observability for live intervention and compliance (Principles 10 & 14).

Unifying Workers AI and AI Gateway into a single AI control plane. Cloudflare consolidates Workers AI and AI Gateway into a single control plane for unified routing, observability, and billing across models and providers. A unified control plane is the operational pattern you need to manage model routing, cost, and provenance across distributed agents and services (Principle 11).

Four AI agents coordinating in real time outperformed Claude Opus 4.8 on enterprise coding tasks. Coral AI Labs shows asynchronous AgentRadio where coordinating agents share discoveries mid-task and double long-horizon code-understanding accuracy versus independent agents. Design agent communication primitives and mid-task disclosure channels—coordination patterns can beat bigger single models for complex, long-running outcomes (Principle 09).