council-of-llms

Multi-model deliberation uses parallel subagents configured on different models to independently analyze a topic, then synthesizes their outputs into a unified verdict with consensus points, disagreements, and action items.

Stress-testing a high-stakes plan involves spawning subagents aligned to distinct analytical lenses like Strategic, Analytical, and Creative, then merging their independent outputs into a decision memo highlighting blind spots and risks.

Configure three independent models in a council config file, provide full topic context inline to the spawned subagents, and the session generates a markdown review file with a weighted verdict and concrete next steps.

Single-model roleplay produces shallow analysis because it lacks genuinely independent perspectives and adequate context control, whereas spawning subagents on different models ensures distinct analytical perspectives are independently generated.

Limitations include context overflow and empty outputs if prompt size limits are exceeded, requiring practical guardrails like inline context preparation, lightContext execution, and timeout handling to reduce these common failures.

Yes, generating a unified decision verdict requires three independently configured models to provide distinct analytical perspectives, ensuring the final synthesis captures genuine consensus and disagreements rather than shallow roleplay.