Project Think: building the next generation of AI agents on Cloudflare launches primitives and a base class for durable, sandboxed AI agents that persist, fork, and scale. This matters because durable execution and sandboxing are the runtime foundations outcome engineers need to make agent behaviors reproducible and safe in production (Principles 07, 09, 06).

Rearchitecting the Workflows control plane for the agentic era reworks Cloudflare’s control plane to support massive agent-driven concurrency, raising instances, throughput, and queued capacity. Outcome engineers should treat control-plane scaling and queuing semantics as a first-class design problem when moving agents from demos to high-throughput services (Principles 09, 04).

Plain — Full‑stack Python framework for humans and agents ships a typed, full-stack Python framework with built-in agent tooling, guardrails, and LLM-friendly docs to power human+agent workflows. Use this as a practical template for end-to-end developer ergonomics: typed contracts, docs, and integrated guardrails reduce friction between engineers and deployed agents (Principles 03, 06, 10).

Curity reinvents IAM with runtime authorization for AI agents introduces Access Intelligence to enforce runtime, intent-bearing tokens and human approvals for risky agent actions. Outcome engineers must design auth and runtime gating into agent flows so agents can act productively without escaping human and legal guardrails (Principles 15, 10).

Anthropic details AI agents accelerating alignment research through ‘weak-to-strong supervision’ uses chained agents to run weak-to-strong supervision, letting weaker models iteratively train stronger ones and speed up alignment experiments. That pattern is a concrete example of agentic pipelines for model improvement and auditability—use it as a template for automated evaluation, data curation, and outcome validation (Principles 09, 16).