One question, several models
Concilium is a method for bringing several AI model families to one difficult question. Instead of asking one assistant for a confident answer, it gives models different roles and lets them explore, challenge and extend one another's work. The human remains responsible for the question, the evidence and the final judgment.
The model families
The council can include Claude, which usually orchestrates the work and keeps the whole problem in view; Codex, which provides an independent OpenAI perspective and can review claims or run its own investigation; Kimi, an optional Moonshot seat that adds another independently trained lineage; and Grok, an optional xAI seat that brings a fourth route through the problem. Their roles can change—the point is to combine models whose strengths and weaknesses do not completely overlap.
Where it helps
Concilium is useful for scientific and scholarly research, hard technical problems, consequential decisions and open questions where the route to an answer is not yet clear. It can search for new directions, test an existing claim, or keep several interpretations alive until evidence separates them. It is most valuable when a normal search or a single model keeps returning the obvious answer, but the problem deserves a deeper one.
Concilium is available through the Claude Code plugin marketplace: add raichominev/concilium, then install concilium@raicho-skills. Its source, documentation and manual installation instructions are public in the Concilium GitHub repository.
Top comments (4)
Multi-model methods get interesting when the disagreement is structured. If three models answer differently, the value is not a vote by itself; it is seeing which assumption each model made. The judge layer should explain the conflict, not just pick the most confident-looking response.
Exactly. The disagreement is the point - it exposes different assumptions. At first, I looked for those differences by hand. Concilium takes it a step further: it automates the exchange, loops ideas between models, measures the proposals against known facts, etc. Sometimes this creates a decisive build up that no single model/family can reach on its own. Thanks for the term — “structured disagreement” describes the idea very well.
That makes sense. The automation is valuable if it preserves the disagreement instead of smoothing it away too early. I would want the final output to show the strongest rejected assumption too, because that is often where the useful learning is.
I actually used you comment to audit how much of it apples and found out that disagreement was somewhat implied, but not in full. Added a real fix - keep and track disagreements explicitly.
I was hoping to get the models quarrel, debate, defend, etc. It did not work for now - they agree too quickly (mostly at the second round).
Midway I added a bunch of viable frontier models - GLM, Qwen, Deepseek, Gemini, except Mistral, might try it at some point also. Instructions published on the git repo, but unreleased officially yet - setting up the whole thing gets quite complicated, irrelevant for most setups I guess.
I am now testing brainstorming as a concept (a spin-off from the disagreement experiment) with the newly added models already generating interesting ideas. I do deep linguistic research on a dead language with limited artifacts to work with available. Currently every concilium test run results in a full-blown development session for the linguistic analyzer - thus concilium method list is not saturated yet.
I was stuck and your contribution matters. Thank you.