Building agentic social media workflows with MCP + a unified publishing API

Most social tools are dashboards. I wanted agents to own the loop: research/plan → generate platform-specific content → schedule or draft → publish → monitor/retry, with human approval gates where it matters. Ended up with a hosted MCP server (one URL paste for Claude/Cursor-style clients) exposing a set of tools for accounts, posts, media, scheduling, analytics, etc., plus a normal REST API and SDK underneath so you can also script it. Supports the usual platforms with per-platform customization and automatic format handling. Drafts stay human-reviewable before they go live. Interesting bits so far: scoped agent access + approvals reduce the “agent does something irreversible” risk, bulk/queue patterns work for evergreen content, and the combination of generation + cross-post + calendar is where the time savings show up. Still fighting platform rate limits, media readiness states, and making failure recovery agent-friendly. Anyone else building agent tools that touch external publishing/ops systems? What patterns are you using for human-in-the-loop or observability? Feedback on the tool surface or architecture welcome.