Dev seeks production architectures for SaaS action-taking AI agents beyond MCP basics

Solid-Instance-3352 · reddit · 2026-09-30

A Reddit developer is asking for real, battle-tested architectures for AI agents that take actions inside a SaaS platform — e.g. turning "book me a hotel for 3 nights" into actual API calls with failure handling.

Beyond the usual MCP/n8n/function-calling tutorials, they want production lessons on: single agent vs. orchestrator + specialized agents; how to expose platform capabilities (REST, MCP servers, tool calling, SDKs); handling auth, permissions and user identity; preventing unauthorized or dangerous actions; multi-step workflows with dependent calls, retries, idempotency and rollback/compensation; where business logic should live; human-in-the-loop confirmation for payments and bookings; production monitoring; and how to scale to hundreds of tools. They explicitly want diagrams, code, GitHub projects and "what worked, what didn't" stories — not another "just use MCP" answer.

Original post →

More from coding & agent

coding & agent channel →