When Does a Second Round Turn Orchestrator-Worker Into an Agent?

A "competitive-landscape research" feature is a textbook orchestrator-worker: an orchestrator call reads the query and emits 2–6 sub-questions, workers research each with a search tool in parallel, and a synthesizer writes the report. Three upgrades are on the table:

• P1: after synthesis, an evaluator call checks the report against a fixed rubric (every sub-question answered, every claim sourced); if it fails, the synthesizer re-runs with the critique, for at most 2 rounds.
• P2: after synthesis, the orchestrator re-reads the worker outputs and the draft, decides what new sub-questions are missing, spawns workers for them, and repeats until it is satisfied or 4 rounds have run.
• P3: identical to P2, but with a hard cap of exactly 1 extra round.

1. For each proposal, decide whether the system stays at Level 1, and reconcile the apparent contradiction that "evaluator-optimizer loops are Level 1" while "an orchestrator that reacts to worker results is Level 2."
2. Describe, in the lesson's own terms, what P2 does to the system's testability and its latency/cost profile relative to the original.
3. The team wants gap detection but wants to stay at Level 1. Design a version that captures most of P2's benefit, and state precisely what it gives up.