Part 1 of 6 · 1 min read
Overview
"Which framework would you use?" is a standard interview question, and the weak answer is a brand. The strong answer names the capability that the problem needs (durable state, approval pauses, cross-provider portability, tracing) and then says which paradigm provides it, with the trade-off. This module gives you that vocabulary and a small experiment showing why one capability, checkpointing, is not the same as another, idempotency.
About the example, and what is real
Real: the paradigm code in code/mini-paradigms (a loop that restarts, a graph that resumes from a checkpoint, an idempotency key), the seeded Monte Carlo, and the 14 tests. Assumed: crashes happen after a step's work and before it is recorded, at a fixed probability per step; the "model" is a script that picks the first unfinished step; token sizes for the handoff comparison (2,000 base context, 500 per turn, a 600-token brief, a 150-token summary). Not real: no framework is imported. This does not measure LangGraph, the Claude Agent SDK or the OpenAI Agents SDK; it shows what the paradigms cost.