9 MCP Resilience Patterns That Keep AI Agents Alive in Production (With Code)

Model Context Protocol (MCP) went from "cool demo protocol" to production infrastructure in about six months. But here's the thing — most tutorials show you the happy path. Connect a server, call a tool, done. Production is different. Production means auth failures at 3 AM, context windows exploding, tools timing out, and agents calling the wrong tool because your descriptions were ambiguous. These are 9 patterns I've battle-tested for keeping MCP-based systems alive in production. Real code. Real problems. Real fixes. Pattern 1: The Circuit Breaker for MCP Tool Calls Your agent calls an MCP tool. The server is down. The agent retries. And retries. And retries. Meanwhile, your context window fills with error messages and your user stares at a spinner. The fix: wrap every MCP tool call in a circuit breaker. import time from dataclasses import dataclass, field from enum import Enum from typing import Any, Callable import asyncio class CircuitState(Enum): CLOSED = "closed" # Normal operat