The frameworks make it look easy. LangChain, CrewAI, Autogen, Model Context Protocol — the docs all show you the happy path. The reality of shipping a production AI agent — with real users, real rate limits, real concurrency, real failure modes — is something else.

In the last two months I shipped four open-source MCP servers covering financial filings, prediction-market data, real-time audio analysis, and stealth web browsing. Two PRs are in code review at Anthropic's official modelcontextprotocol/servers repo. The packages have crossed 5,000 NPM downloads. The same architectural patterns and failure modes apply whether you're building on MCP, CrewAI, or rolling your own agent loop.

This is the talk I wish someone had given me on day one. We'll cover the three production bug shapes that nearly killed each server (the asyncio.run inside async runtimes that crashes silently, the Promise.all vs Promise.allSettled choice that determines whether one failed source poisons your whole response, the print() statement that corrupts stdio JSON-RPC). The token-economics decisions that turn an expensive agent into a cheap one (representation choice, response shape, cache strategy). And the README-as-architecture principle — why what you say about what your agent isn't matters more than what you say about what it is.

If you're about to ship your first production agent or you're a few weeks into one and wondering why it keeps falling over, this talk has the lessons that took me two months of compressed building to learn.