framework-selectionlisted

# Framework Selection The single best piece of advice from Anthropic's "Building Effective Agents," and it's worth a million dollars: **"Start by using LLM APIs directly: many patterns can be implemented in a few lines of code. If you do use a framework, ensure you understand the underlying code."** HumanLayer's Dex Horthy independently observed: in production customer-facing agents, most teams roll the stack themselves rather than committing to a heavy framework. The framework is a convenience, not a requirement. That said, when you do choose, do it by constraint, not by hype. ## Step 0: do you need a framework at all? For v1 / proof-of-concept, often **no**. The model provider's SDK plus your own code handles the 5 patterns cleanly: - Prompt chaining = a Python function calling the LLM twice - Routing = a switch statement after a classifier call - Parallelization = `asyncio.gather` - Orchestrator-workers = a function spawning sub-tasks - Evaluator-optimizer = a while loop If you can express your design in 100 lines of vanilla code, do that. You'll understand it better, debug it faster, and have fewer dependencies. Frameworks earn their place when you have: - Long-running workflows that need durable execution - Multi-agent coordination with complex state - Need for production observability with deep integration - A team that benefits from shared vocabulary and structure - Cross-cutting concerns (HITL, checkpointing) you'd otherwise rebuild ## The decision matrix (co