Agreement is a bug. I forced 11 Claude Code agents to disagree. (2026). [microcontent]. Nyk. X, March 19, 2026. Retrieved 2026-03-19 from: https://x.com/nyk_builderz/status/2034492549180625316
council-of-high-intelligence (2026). [software]. Nyk (0xNyk). GitHub, CC0/MIT license. Retrieved 2026-03-30 from: https://github.com/0xNyk/council-of-high-intelligence
Nyk builds a Claude Code skill that externalizes disagreement into a multi-agent council of historical thinkers and contemporary technologists. The X thread describes the founding architecture: 11 personas in 6 polarity pairs with a 3-round protocol. The repository expands this to 18 members, 13 polarity pairs, 20 pre-built triads, three deliberation modes (full, quick, duo), and multi-provider auto-routing that separates polarity pairs across different language models. Each persona declares its analytical method and blind spots. The verdict leads with what the council does not know rather than with a confident recommendation.
Historical-thinker anchoring produces tradition-specific reasoning: Named historical figures activate richer behavioral patterns than abstract role descriptions because the model has encoded each figure’s actual argumentative style. Socrates activates the elenctic method specifically, not generic questioning. Feynman activates first-principles stubbornness and distrust of unexplained complexity. The tradeoff is stereotype risk: the model’s representation of each figure is filtered through popular reception, which may overrepresent caricature and underrepresent scholarly nuance.
Polarity pairs form a graph, not a partition: Members appear in multiple pairs (Torvalds in three, Karpathy in three, Ada in three). The same persona produces different analytical output depending on which opposition it faces, because different aspects of its intellectual tradition are activated by different counterweights. This means persona behavior is context-dependent rather than fixed – a structural insight about multi-agent persona design.
The Problem Restate Gate catches framing errors before analysis: Every council member restates the question and provides an alternative framing before analysis begins. Divergent restatements surface that the question itself is poorly framed. This meta-cognitive step operates above the analytical layer and is absent from Lehmann’s architecture, which begins directly with independent analysis.
Dissent quotas mechanically prevent premature convergence: If more than 70% of members agree too early, two members are forced to steelman the opposing view. This treats convergence as a failure signal rather than a success signal. The forced steelman is harder than mere disagreement and produces qualitatively different output.
Multi-provider routing sidesteps the intra-model diversity question: Rather than debating whether single-model personas produce genuine reasoning diversity, the council distributes polarity pairs across Claude, OpenAI, Gemini, and Ollama. Different models for the most important disagreements; same model acceptable for agreement.
Triads are sequential reasoning pipelines, not just perspective panels: The architecture triad (Aristotle, Ada, Feynman) follows a progression: classify, formalize, simplicity-test. Each member’s output feeds the next. This sequential structure differs from the oppositional structure of polarity pairs. Pairs produce tension; triads produce analytical progression.
Verdicts lead with uncertainty rather than confidence: The synthesis step produces “Unresolved Questions and Recommended Next Steps” before any recommendation. This inverts the typical multi-agent synthesis pattern, which collapses uncertainty into a single action. The human is treated as the chairman.
“The biggest failures weren’t ‘wrong answers.’ They were blind spots from a single perspective.” (X thread – problem statement)
“LLMs don’t truly think in parallel. They simulate one coherent viewpoint per generation. So I externalized the disagreement layer.” (X thread – architectural rationale)
“Each agent declares its analytical method, what it sees that others miss, and – critically – what it tends to miss.” (X thread – persona specification)
“A single LLM gives you one reasoning path dressed up as confidence. Ask it a hard question and you get a fluent, structured, wrong answer.” (Repository README – extended problem statement)
“If 3 members restate your question differently, the question was the problem.” (Repository README – Problem Restate Gate rationale)
“What the council doesn’t know matters more than where it agrees.” (Repository README – uncertainty-led verdict rationale)
“If >70% agree too early, two members are forced to steelman the opposing view.” (Repository README – dissent quota mechanism)
The council represents the most architecturally complete implementation of historical-thinker-anchored multi-agent deliberation in the practitioner space. Its polarity pair structure, Problem Restate Gate, dissent quotas, and uncertainty-led verdicts each contribute distinct mechanisms not present in Lehmann’s LLM Council or other multi-agent deliberation systems. The repository’s expansion from 11 to 18 members and from 6 to 20 triads shows active evolution. With 105 GitHub stars and CC0/MIT licensing, the implementation is available for direct reuse. The X thread received 36,884 views and 245 bookmarks, indicating practitioner interest. For the garden, this citation is load-bearing for the Structured Disagreement Through Persona Review pattern and directly informs several open questions in [[Persona Design Choices Across Analytical Cultural and Professional Axes]]↑.