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. Durable primitives change how you model agent lifecycles and state, letting teams reason about long-running tasks and operability instead of ad-hoc scripts (Principles 07, 09).

The next evolution of the Agents SDK delivers a model‑native harness and native sandbox execution, enabling developers to build safe, long‑horizon agents. Native sandboxing and an in‑distribution test harness give outcome engineers a repeatable CI-like environment for validating behavior and reducing drift once agents run in production (Principles 07, 14).

Inside VAKRA: Reasoning, Tool Use, and Failure Modes of Agents analyzes agents on executable, tool‑grounded enterprise tasks and exposes pervasive multi‑step reasoning failure modes. Benchmarks that exercise tool use and real workflows let you surface brittle chains early and design observability and validation into agent pipelines (Principles 16, 14).

Salesforce launches Headless 360 to support agent-first enterprise workflows introduces a programmable control layer that maps CRM data and workflows into an agent‑friendly API. Turning CRM into a composable control plane shortens integration work and forces explicit data contracts, which is essential when agents must act deterministically and auditable across business processes (Principles 09, 06, 10).

MuleSoft Agent Fabric adds new ways to keep AI agents in line extends Agent Fabric with deterministic routing and centralized LLM governance to rein in agent sprawl and control costs. Centralized routing and governance are the operational levers outcome engineers need to enforce safety, cost controls, and policy across fleets of autonomous actors (Principles 09, 10).