Agents as Infrastructure: runtimes, tool-calls, and skill authoring

Everything we launched during Agents Week rolls out Cloudflare’s agent runtime, an ADLC, Zero Trust for agents, and an Agentic Internet vision. Outcome engineers now have a concrete stack to build and deploy autonomous apps — treat these primitives as platform dependencies, not experiments.

Mistral patent: Code implemented tool calls patents LLM-generated code blocks that pause for client-executed tool calls, enabling sandboxed, resumable tool orchestration. That changes how you design tool-call contracts, resumability, and safety boundaries for multi-step agent workflows — plan for explicit pause/resume semantics.

Can Agents Use a Computer Yet? We’ve Got the Data shows empirical evidence that computer-using agents are production-ready for standardized back-office workflows and that value is shifting from raw navigation to context, validation, and process knowledge. Outcome engineering should re-center on context engineering, validation pipelines, and capturing process expertise rather than only improving LLM prompts.

Claude Code for Normal People: Skills, Voice Mode, and Collaborating with AI explains running a service with Claude Code using skill files, voice guides, and automated pipelines as first-class artifacts. Treat skill files and voice workflows as productized interfaces for agents — they’re the unit of delivery for team coordination and reproducibility.

Claude’s Record-a-Skill cut my research from hours to 30 minutes — but the magic has limits turns narrated screen recordings into reusable Claude automations that drastically speed work while exposing fragility, high cost, and slow runs. This is a reminder that skill authoring alone isn’t enough: build validation, cost controls, and observability into skill pipelines to catch drift and brittle automation.