Build the Decision Process Around Stochastic Consensus

The council's reasoning method is described as stochastic consensus. Instead of asking one model for one direct answer, the system creates multiple AI agents with different configurations. Each agent develops an independent perspective, and the agents then argue, debate, or review the alternatives before the system produces a final consensus.

The important shift is architectural. The task is no longer a single prompt followed by a single response. It becomes a staged process in which several analyses are generated and then reconciled.

The council workflow contains four essential stages:

1. Multiple agents are spawned for the same question.
2. Each agent receives its own prompt or configuration and develops an independent view.
3. The agents' outputs are compared through debate or review.
4. A final decision stage produces the council's answer.

Independence is important because the system is trying to avoid one immediate line of reasoning controlling the entire result. Separate agent prompts create several candidate analyses. Debate and review then expose disagreements that would remain hidden in a single-response workflow.

The final answer is therefore downstream of multiple perspectives. It is not simply the first response generated by the underlying model.

A direct question can produce an answer before the model has been forced to examine competing interpretations. Stochastic consensus adds deliberate friction. Each agent must contribute a perspective, and the final stage must account for those perspectives before making a call.

This process is presented as consistently outperforming the simple approach of asking the question directly. Its value comes from the combination of independent analysis and synthesis, not merely from generating more text.

Once the skill is installed, invoking the council causes the system to spawn its sub-agents automatically. The user does not manually open separate chats for each adviser. The skill coordinates the agents, their prompts, and the final response.

During execution, the user can open individual agent traces and inspect the prompt and response associated with each agent. This makes the process observable. Rather than seeing only the final recommendation, the user can examine which perspectives were created, which agents have completed their work, and how the council reached its conclusion.

This inspection step is useful when the decision is important because it exposes the individual analyses that precede the final recommendation.

Stochastic consensus introduces more moving parts than a direct prompt. It requires a prepared council skill, multiple agent calls, and a final review stage. This operating procedure treats the orchestration code as a prepared component rather than explaining how to build it from scratch.

The method also has two different forms of diversity: prompt diversity and model diversity. A council can use several roles while still running every role on the same underlying model. That distinction becomes critical when evaluating the default and cross-vendor modes.

Stochastic consensus turns one question into a structured multi-agent deliberation: independent analyses first, debate or review second, and a final decision last. The council's value comes from this sequence, not from asking the same question repeatedly without coordination.