I'm building with MCP and starting to run agents with more autonomy — multiple tools, longer sessions, less human oversight. The failure modes I keep hitting aren't the model being wrong, they're the agent making too many calls, retrying things that won't work, or calling tools it shouldn't need for the task. Right now I'm just setting hard iteration limits in my agent code, which feels like a blunt instrument. What are other people doing here — especially anyone running MCP agents in anything close to production?