Your Agent Framework Is the Integration Database Anti-Pattern in Disguise: What AutoGen, LangGraph, and CrewAI Actually Are
A new deep-dive from Build Simple has done something most framework evaluations avoid: instead of comparing AutoGen, LangGraph, CrewAI, ByteDance's DeerFlow, and Anthropic's published agent patterns on features and pricing, it reverse-engineers each one against the classic architecture pattern literature — and identifies exactly which 40-year-old enterprise integration pattern each framework actually implements. The findings are pointed. AutoGen is the integration database anti-pattern: agents broadcast through a GroupChatManager that reads every message, and token costs compound to 4x–8x a direct call by round 20. LangGraph is the Saga pattern — distributed transaction coordination with compensating transactions. CrewAI is pub/sub with the LLM as a runtime topic filter. Anthropic's composable primitives are scatter-gather, pipeline, and orchestrator-worker. Chapter by chapter, it's the Enterprise Integration Patterns book.
The core argument isn't that frameworks use old patterns — old patterns are battle-tested precisely because they've been used and debugged at scale. The problem is renaming them. Engineers who don't know LangGraph is a Saga implementation can't consult the Saga literature on partial failure handling. Engineers who don't recognize AutoGen as the integration database anti-pattern won't see the token explosion coming until it hits production. The vocabulary gap costs teams real money and real debugging time.
The second half of the piece examines what the author calls the routing-logic-in-prompts problem: when a framework's coordination logic lives in a system prompt rather than code, it can't be unit tested, it can't be diffed, and it silently drifts when the underlying model updates. The Anthropic retrospective on two billion workflows — "start with 100% human review, work down to 50%" — is cited as deployment best practice that, read another way, is an admission that routing logic in prompts is fundamentally untestable. The article doesn't argue against using frameworks; it argues for knowing what you're actually using — so you can ask the right questions about failure modes before you're in production trying to answer them.