Lehmann adapts Karpathy’s LLM Council — originally a multi-model polling system — into a Claude Code skill that spawns five sub-agents with adversarial thinking roles. Each advisor responds independently, responses are anonymized, five reviewers evaluate the full field, and a chairman model synthesizes the final recommendation. The peer review round, particularly the question “what did all five miss?”, produces the highest-value output by detecting gaps in the negative space between perspectives. A product-format case study shows the architecture catching a reinforcing relationship between arguments that no individual advisor identified.
Role-constrained diversity replaces model diversity: Karpathy’s council polls different models; Lehmann uses five thinking-style roles within a single model. Whether role-constrained prompting produces genuine reasoning diversity or the appearance of diversity from one probability distribution remains an open question with direct implications for persona architecture.
Anonymization prevents role-based evaluation bias: After the five advisors respond, the skill shuffles which advisor maps to which letter before peer review. This forces reviewers to evaluate content rather than deferring to or dismissing responses based on the role that produced them. The estate’s persona review currently lacks this step.
The peer review round, not the five-way divergence, creates the most value: Lehmann identifies the comparison step as where blind spots surface. Five perspectives side by side make the gap between them visible — and more importantly, make the space where all five are silent into a detectable signal.
Gap detection operates on absence, not presence: The “what did all five miss?” question detects what the field of responses fails to cover. This differs from ensemble averaging (cancels noise) and adversarial debate (tests claims). It is a distinct multi-agent mechanism that deserves its own pattern name.
Emergent argument interactions span role boundaries: In the case study, the Contrarian’s completion-rate argument and the First Principles speed argument reinforce each other in a way neither saw. This reinforcing combination — not either argument alone — is the strategic insight. Emergence requires a structured comparison step that individual agents cannot perform.
Five adversarial roles cover five decision-making blind spots: Contrarian (risk blindness), First Principles (assumption lock-in), Expansionist (scope myopia), Outsider (expertise bias), Executor (implementation gap). The set lacks roles for ethics, long-term consequences, and systemic effects — suggesting it is a minimal covering set for business decisions, not a universal one.
Chairman synthesis trades human authority for decision speed: The architecture ends with a model synthesizing a “clear recommendation and one concrete next step.” This reintroduces sycophancy risk at the synthesis layer and contrasts with the estate’s pattern where the human decides after reviewing structured disagreement.
Sycophancy framed specifically as framing sensitivity: The article defines the sycophancy problem as the model picking up on “your assumptions, your framing, your emotional lean.” This is narrower than the full sycophancy problem but has a clear architectural solution: force multiple framings simultaneously.
“Claude just tells you what you want to hear. Every time. You can’t trust it.” (Opening — problem statement)
“I rebuilt it to work entirely inside Claude using sub-agents with different thinking styles instead of different models.” (Architecture — the key substitution)
“After all 5 advisors respond, the skill anonymizes everything. Shuffles which advisor maps to which letter. The reviewers don’t know who said what.” (Peer review — anonymization step)
“That last question is the most valuable one. Every time I’ve run the council, the peer review round catches something no individual advisor saw.” (Peer review — gap detection claim)
“When you read 5 perspectives side by side, the gap between them reveals what nobody thought to mention.” (Peer review — negative space mechanism)
“The Contrarian’s completion rate argument and the First Principles Thinker’s speed argument actually reinforce each other in a way neither advisor saw alone.” (Case study — emergent interaction)
This article provides the most concrete practitioner implementation of structured multi-agent disagreement for decision quality. Its direct connection to the estate’s Structured Disagreement Through Persona Review pattern makes it load-bearing for the persona strategy discussion: it demonstrates anonymization, gap-detection naming, and the role-vs-model diversity question that the estate’s architecture must address. The 198,000+ views and 2,000+ bookmarks indicate significant practitioner uptake of the council pattern.