Part 1 of 6 · 1 min read
Overview
Frameworks present shapes as classes (a supervisor, a swarm, a group chat). The interview question is what each shape costs, what it can lose, and when to use which. This module treats a shape as a small program with calls, tokens, a critical path and a failure mode, and measures them on one task.
About the example, and what is real
Everything runs offline in code/mini-shapes. Real: the five control flows (including the handoff state machine with a hop limit and the critic loop with bounded rounds), the Monte Carlo over 20,000 seeded tickets, the call, token and critical-path accounting, and the 16 tests. Assumed: the agents are probabilistic stand-ins: specialist accuracy 0.85 per part, generalist 0.70, router 0.90, decomposition 0.90 per domain, a misrouted specialist notices 60 percent of the time, a critic detects 70 percent of wrong drafts and rejects 10 percent of correct ones. These numbers are mine; what the experiment shows is how each shape transforms them. No model is called.