Lab: Decide whether to split, and prove it

Module 11 · Lab · About 75 minutes · Needs: the Module 3 lab's tools.py, and lessons 1, 2 and 3 (one agent first and return_workflow; orchestrate, merge and fan-out by order; TaskSpec, compress, the order lock and the cost ledger)

What you are building

It is Thursday 4 June 2026, week 8. Dev's whiteboard still shows three agents, and the three lessons have tested every shape on it. This lab turns their answers into code and a written decision. returns_desk/team.py holds the mechanics any split must use: a task spec enforced in code, a subagent loop that charges every model call to an agent, one budget per request, compressed results with a trace id, a read-only brief of the queue, and one writer per order. returns_desk/decision.py measures one agent against the split on R-1001 and writes docs/multi-agent-decision.md from the numbers. Both import Module 3's tools.py and never replace it. Everything at BrightCart is simulated.