The Big Shift in MCP: Why AI Guides Will Replace API Wrappers (www.tigerdata.com)

AI guide is a proposed shift in the MCP (model-connected platform) architecture: instead of handing models neutral API wrappers that expose capability, developers should embed expert judgment into opinionated “Guides.” Matvey Arye announced pg-aiguide, a reasoning layer for Postgres that sits between agents and the DB and enforces structured workflows, best-practice patterns, and semantic grounding. The motivation: models can validly execute API calls (valid JSON, successful DB write) while making poor engineering choices at machine speed—unindexed foreign keys, wrong timestamp types, queries that force sequential scans—because wrappers define what’s possible, not what’s wise. Recent advances like Code Mode / Code Execution make agents faster but amplify this judgment deficit. Technically, a Guide bundles curated “skills” (stepwise workflows), retrieval-grounded reasoning (semantic search of authoritative docs/man pages), and enforced decision paths so an agent is pushed through deliberate design choices rather than left to freeform commands. This pattern—structured, verifiable workflows—mirrors trends across Claude Skills, ReAct-style flows, and reasoning chains from Cursor/Windsurf, but Guides aim for portability via MCP tools so any model can use the same standards. Implications: safer, production-ready agents that reduce machine-speed technical debt, cross-model interoperability of guardrails, and a new open-development role for domain experts to encode institutional knowledge into runtime tooling.