Multi-Agent Research System

What this scenario tests

The orchestration scenario. It exercises task decomposition, what a coordinator passes to a subagent and gets back, and how a claim stays attached to its source through several hands. The reliability half matters as much as the architecture half: a research system that loses provenance produces confident output nobody can check, which is the failure this scenario is built around.

Traps it is built to catch

• ▸Parallelising sequential work. Subagents run in parallel profitably when their subtasks are genuinely independent; splitting interdependent steps just multiplies the handoffs where state gets dropped.
• ▸Passing the whole conversation to a subagent. Explicit, structured context is the point — a subagent inheriting everything inherits the coordinator's context problems too.
• ▸Treating provenance as a formatting step at the end. By the time synthesis runs, an unattributed claim cannot be traced back; the mapping has to survive each handoff.