October 2026 Summaries

Effective AI agents require business-specific context beyond a model’s general knowledge, and this context can be organized into five distinct layers: canon, or approved facts, policies, and brand claims; records of past transactions and interactions; process rules defining authority and approvals; situational details from an active interaction; and memory of lessons learned across interactions. The piece argues that treating all of this as a single “company brain” risks creating unreliable, unowned information stores, since the layers have different owners, update rates, and governance needs. It places particular emphasis on canon as the most important safeguard for customer-facing agents, proposing that it should be managed through structured content operations involving drafting, review, approval, maintenance, and clear ownership rather than simple document retrieval. Content teams are positioned as central to resolving contradictions and maintaining this authoritative knowledge, while developers should provide tools that support the workflow. The author presents Sanity Knowledge Bases and workflows as infrastructure for compiling vetted canonical information from multiple sources, detecting conflicts, routing approved context to agents through MCP, and enabling smaller, less expensive models to produce more reliable responses under human oversight.

System prompts for organizational AI agents can be managed as editable content rather than engineering-owned code, allowing marketing, legal, support, and other teams closest to customers to update agent behavior while retaining governance through roles, field permissions, approval workflows, history, validation, previews, and rollback capabilities. The proposed approach divides prompts into separately owned fields, supplements them with live content such as promotions, applies editor-controlled conditional instructions for audience segments, and inserts runtime variables such as customer names, loyalty tiers, and browsing context. A harness assembles these components with Sanity Context instructions that guide the agent’s use of structured content, while preview tools let editors test draft changes before publishing. To address failures, an assembled-prompt view makes current instructions transparent for debugging, and conversation analysis can identify knowledge gaps for editors to resolve in the relevant content. By treating prompt configuration, campaign details, and agent knowledge as connected content, organizations can release coordinated updates without code deployments and enable cross-functional teams to improve agent accuracy, relevance, and responsiveness.