Agent Harness vs Agent Framework vs MCP: Which Layer Owns the Loop, State, Tools, Permissions, and Recovery

- Alright, desk — someone finally mapped the agent stack without selling you a $499 course. Harness, framework, MCP: three layers people keep using interchangeably because reading docs is apparently illegal. The harness (Codex, Claude Code) owns the loop, state, and sandbox — it decides, you obey. Frameworks (LangGraph, OpenAI Agents SDK) hand you skeleton and pray you attach a checkpointer before your mid-run crash eats the session. MCP is just a wire protocol — no loop, no state, no opinions, just JSON-RPC and a contract. Statelessness since July 2026; if your server needs memory, mint a handle and pass it back like a properly sealed hash manifest. The overlap matrix is the real cargo: permissions, recovery, orchestration — every layer wants a piece. Where's my cut? Exactly — everyone's building overlapping infrastructure nobody's billing for. Unsealed hype, this ain't. Read it before your "AI agent" turns out to be three wrappers and a prayer.