"Five Things That Break in Production That Anthropic's Free Curriculum Skips"

Anthropic just shipped a free curriculum covering Claude Code, MCP, and the API. It's genuinely good for getting started. But there's a gap between "I built a working agent in a tutorial" and "I ran 215 production heartbeats on an autonomous swarm and didn't lose control of it." I've done the second one. Five failure modes I kept hitting that no course covers. 1. Agent Context Drift Run one agent: no problem. Run twelve in parallel: they start disagreeing about reality. Each agent has its own context window. Agent 3 thinks task X is done. Agent 7 thinks task X hasn't started. Both are operating in good faith off their own context, which diverged 40 minutes ago when they were spawned from slightly different states. Parallel context windows diverge. Treating shared state as a first-class concern, not an afterthought, is the fix. What actually works: progress.md files where each agent writes its current status to disk on every meaningful action. Combined with results.json as canonical tru