Agent Council Orchestrator

Agent Council Orchestrator answers ordinary questions the way any assistant would, then reports the measured truth about how the answer was produced. A router reads the question, judges its difficulty and spends accordingly: simple questions go to a single model and stream back in roughly two seconds, while harder ones are split across specialists running in parallel and merged into one reply delivered as prose rather than a findings table. The specialists are derived from the question itself rather than a generic template, so asking about ledger locking produces locking-strategy and ledger-architecture agents. A second surface handles heavier work: the user states a goal, the system asks clarifying questions and permission, designs an agent team as a visual graph, shows it for approval, executes parallel-first, streams live state and can replay any past run with a provably identical result. Every edge in the graph carries a JSON Schema, so a wrong shape is rejected and the exact validator error is returned to the producing agent for retry. The whole system runs at zero cost on free tiers.