Agent clouds, iframe tooling, and cheaper frontier inference

Welcome to Agents Week — Cloudflare kicks off Agents Week to define an agent-native Agent Cloud, publishing primitives and developer-lifecycle guidance around security, context engineering, and the agentic web. Outcome engineers get a concrete vendor framing for agent infrastructure and lifecycle patterns — this pushes Orchestration toward production-grade primitives you can build against (Principle 09).

datasette-apps 0.2a0 adds invisible-iframe debugging and app-listing tools so agents can test and manage apps programmatically. This gives you a lightweight, inspectable way to instrument agent interactions with web apps and improve observability and debugging during agent-driven workflows (Principles 03 & 06).

GM building in-vehicle AI assistant because Gemini can’t access car data reports GM is building an in-vehicle AI that directly accesses OnStar and vehicle telemetry because Gemini can’t access that data. If your agents need to control hardware or private telemetry, you must design secure, auditable data bridges and clear consent/permission layers — a practical call for Legible Landscapes and Graph access controls (Principles 06 & 11).

Explorative modeling: Train on the best of K guesses presents XM, a training method that adds an exploration axis to generative training and improves sample, FLOP, and parameter efficiency across images, video, and language. Outcome engineers can use this to shrink training budgets and iterate models faster, altering trade-offs in model selection, validation, and deployment (Principles 12 & 16).

Running Kimi K3 on MI355X at Better Performance per Dollar Than B300 shows MI355X running Kimi K3 with superior performance-per-dollar compared to B300/B200, enabling cost-effective frontier-model inference. That changes deployment economics: you can push heavier model workloads to edge or on-prem hardware and re-evaluate latency, cost, and gating decisions in agent orchestration (Principles 07 & 15).