Agents have their own computers with Sandboxes GA. Cloudflare launches Sandboxes GA with secure credential injection, snapshots, PTY and Active CPU pricing to run untrusted agent workloads at scale. This gives engineers safe, snapshotable, credentialed runtimes for agent experiments and production, reducing blast radius and operational risk (Principle 07).

Hands-on with the Google Agent Development Kit. Google’s ADK makes building, deploying, and orchestrating modular AI agents easier across languages and runtimes, with human-in-the-loop controls and agent-to-agent support. Outcome engineers can prototype multi-agent workflows faster and inherit runtime controls for governance and coordination (Principle 09).

Google Cloud introduces QueryData to help AI agents create reliable database queries. QueryData gives agents deterministic, near-100% accurate SQL queries by encoding schema-aware context and validation into agent workflows. This removes a major source of hallucination and flaky data access, making agentic data operations auditable and safer for production (Principles 02, 06).

Agentic coding at enterprise scale demands spec-driven development. The article argues spec-driven development makes autonomous agents verifiable and speeds enterprise delivery with continuous, test-backed agentic coding. For outcome engineers, formal specs, test harnesses, and CI provide the controls and observability needed to ship reliable agentic features (Principle 14).

Enterprises power agentic workflows in Cloudflare Agent Cloud with OpenAI. Cloudflare enables running OpenAI models like GPT-5.4 inside Agent Cloud to support production-ready, edge-scalable agent workflows. Engineers can deploy models closer to users, combine edge durable state for per-agent contexts, and better control latency, cost, and security for agentic products (Principles 07, 06).