Multi-agent research sessions

The orchestrator is code. It gives the floor in a fixed rotation, for the number of rounds set, it counts the tokens and it stops the session at the budget. It makes no model call of its own, so it adds zero tokens to the relay bill. One call is in flight at a time; the next one starts once the first has answered.

The three agents marked default below open a session when none is named: the quant researcher formulates the edge and the protocol, the devil's advocate attacks it, the validator judges. You can add the others with an @ in the command, or from the chips above the topic.

Each agent carries a set of tools it may call and a set of refusals written in its persona: the researcher hands a result to the validator instead of declaring it proven; the developer writes the code the researcher describes. A turn is an ordinary assistant call with that persona active and the transcript of the earlier turns attached, so the persona, the coded refusals and the tagged thread built for a single agent apply as is.

An agent states an idea in a tagged block. The verdict rule reads the evidence the idea cites: a run present in the app history, or a note supplied to the session. With such evidence the idea is marked supported; without it, the idea stays proposed and the journal records why. The agents propose and the engine judges, which is the same boundary the assistant keeps on a single conversation.

A hand-off is explicit. An agent that ends its turn with a request to a named colleague gives that colleague the floor at once, once, and the rotation resumes afterwards. The session page draws the ideas, the thread and a graph of who cited what.