Build agents, not pipelines. Sean Goedecke argues for agentic systems over static pipelines to handle complex, iterative tasks, shifting how teams structure solutions. Outcome engineers should favor agent architectures and context-engineering patterns when predictability from pipelines gives way to flexible problem-solving — Principle 06/11.

How we contain Claude across products. Anthropic lays out strict process, VM, filesystem, and egress sandboxes across Claude products to prevent data exfiltration and limit runtime risk. This matters because production agents require baked-in runtime containment and egress controls to protect secrets, meet compliance, and make deployments auditable — Principle 07/14.

What Is an AI Prompt Injection Attack? The Hidden Threat Hijacking Your Chatbots. The guide catalogs prompt-injection techniques that silently hijack chatbots and suggests defenses like input sanitization and human checkpoints. Treat prompt injection as a first-class security failure mode and add validation, provenance, and gating to agent flows to keep outputs trustworthy — Principle 14/15.

How global logistics got me over my fear of personal agents. Matt Webb builds a Claude Code skill to automate customs worksheets for mass shipping, turning a personal-agent experiment into an operational win. Use this as a template for shipping small, task-specific agent skills that augment humans and integrate into existing processes — Principle 03.

Ernst & Young published cybersecurity report full of hallucinations. An investigation finds widespread fabricated citations in an EY cyber report, exposing how hallucinations corrupt research and public trust. Outcome engineers must enforce ground-truth checks, citation validation, and audit trails for agent outputs to prevent false artifacts from escaping into downstream systems — Principle 02/13/16.