When One AI Model Conducts Others: Where a Fugu-Type Chain Can Slip, and a Warning Number to Catch It Early

Systems such as Sakana AI's Fugu place one coordinating model in front of several other AI models. The coordinator chooses which model works on each step, asks another model to check the work, and merges the results into one answer. This concept note reads the published papers behind such systems (Trinity and the Conductor) and points to seven places where a chain of this kind can slip: the stop decision that rests on one model's judgment, a very delicate reading signal, a coordinator that can call itself, fixed limits that do not show drift, one worker's slip passed to the next step, a pool of models that changes, and a merged answer that carries no sign of how firm it is. The note then suggests a simple, training-free addition: a warning number that measures how slowly the small everyday wobbles of a few signals settle back, read over a series of runs. It is inversely proportional to the size of the leading eigenvalue, so a rising number shows the chain approaching a tipping point. Three calm states follow: Steady, Watch and Look closer. A computer-decided shunting action then acts without waiting for a person. It moves the doubtful runs to a simpler path or an extra check, while all other runs carry on and nothing is stopped. The note describes an idea only. It has not been tested on a real system. Patent pending. Suggested keywords: multi-agent AI; LLM orchestration; early warning; critical slowing down; AI safety; coordinator model; drift