Llm council architecture

LLM Council Architecture

Purpose

Designs the control-flow skeleton of a multi-model council: how many stages, what each stage sees, when to fan out vs. run sequentially vs. debate, when to stop, and how to keep the pipeline alive when a member or the chairman fails. This is the pipeline shape — not which models sit on the council, how their votes get scored, what the prompts say word-for-word, or what it costs to run.

When to use

• "Design/build an LLM council" for any task.
• "Multi-model deliberation architecture" / "structure a debate, review, or synthesis pipeline."
• Terse asks: "council architecture?"
• Requests that describe the shape without naming it: "I want three models to answer and one to merge the answers, how do I structure it?", "one model drafts, two critique it, a fourth writes the final version."
• Any question about debate rounds, stopping rules, or what happens when a council member or chairman fails mid-run.

When NOT to use (sibling skills own these)

• Whether to use a council at all vs. a single model → llm-council-when.
• Which models to put on the council → llm-council-members.
• How to score/aggregate rankings into a final answer → llm-council-aggregation.
• Exact wording of the prompts each stage sees → llm-council-prompts.
• Cost/token-budget tuning → llm-council-cost.
• Deeper failure-mode taxonomy beyond degradation-by-design → llm-council-failure-modes.
• Headless coding-CLI councils (e.g. running multiple CLI agents) → llm-council-harness.
• Generic LangGraph app questions with no council/deliberation shape, or a single-agent workflow → out of scope entirely.

If the request is really about one of these, name the sibling and hand off instead of designing the pipeline yourself.

The reference pattern

Source: karpathy/llm-council (Nov 2025, github.com/karpathy/llm-council).

1. Stage 1 — parallel fan-out. The identical prompt goes to N models at once, with no cross-visibility between them. Wall-clock cost is the slowest member, not the sum. Independence is the point: it decorrelates errors and prevents anchoring on whichever answer arrives first.
2. Stage 2 — anonymized peer review. Responses are relabeled "Response A/B/C..." before any member ranks the others on accuracy and insight. Anonymization is the single load-bearing design choice. Revealing real model identities during deliberation measurably increased behavioral convergence (cosine similarity 0.56 → 0.77, p = 0.001; arxiv.org/abs/2604.00026) — anonymity is what blocks brand bias, deference to a "flagship" name, and self-preference. Karpathy's original does not exclude self-votes: each model also ranks its own (anonymized) answer, and his own observation is that models are "surprisingly willing to select another LLM's response as superior to their own." (The derivative llm-council-core PyPI package adds self-vote exclusion — a deliberate deviation, not a bug, if you choose it.)
3. Stage 3 — chairman synthesis. A chairman/synthesizer model compiles the rankings and responses into one final answer. This is the known weak point (github.com/karpathy/llm-council/issues/3, "Chairman over-influence"): a dominant chairman re-homogenizes whatever diversity the panel produced. Two variants: chairman as a regular council member (cheap, reuses a call) vs. a separate model that never generates its own answer (avoids self-bias in synthesis). A stricter variant keeps the judge blind to the raw user input entirely, showing it only the structured stage outputs.

Cost: a full council run is 2N + 1 model calls for N members (N draft + N rank + 1 synthesize). Cost/latency tuning itself is llm-council-cost.

Topology by task shape

Pick the cheapest topology that fits the task — see references/topologies.md for the full evidence and citations.

Topology	Use when	Cost/latency
Parallel fan-out	Breadth, blind-spot detection, latency-sensitive	Near single-call latency
Sequential refinement	Auditable pipeline with real stage dependencies	Latency compounds; add a verification gate or early errors propagate uncaught
Debate/deliberation	Genuinely contested, high-stakes calls only	Most expensive and most fragile — reach for it last

Default to fan-out or sequential; debate is a deliberate, expensive choice, not the default council shape. See references/topologies.md for the evidence backing this ordering (majority-voting-explains-most-gains, adversarial-member-accuracy-collapse, vendor single-agent-first guidance).

Round caps and stopping rules

• Cap debate at ≤ 3 rounds — sycophancy compounds with additional rounds.
• Prefer a statistical stopping rule over a fixed round count. A Wald-SPRT compute governor converged in an average of 1.01 rounds / 4.06 calls vs. 15 calls for fixed-5-round debate, at 97.0% vs. 99.0% GSM8K accuracy — a reported 3.7x call cut for a 2-point accuracy cost (arxiv.org/abs/2605.19193, author-reported figures).
• Early agreement is not a correctness signal. Consensus is cheap to reach and cheap to trust wrongly because member errors are correlated. Treat fast agreement on a high-stakes run as a flag to sample once more or route to a human — not as confirmation. See references/topologies.md#round-caps-and-stopping-rules for the full argument and numbers.

Structured interfaces between stages

• Pass structured JSON between stages, not free prose — a parseable ranking contract is what makes Stage 2 → Stage 3 machine-checkable.
• Instruct the synthesizer as a strict reporter: it may compile and weigh what members said, but must introduce no new facts absent from member outputs.
• Route disagreements by type, not uniformly:
  • Factual → require evidence/tool retrieval before resolving.
  • Interpretive → surface the competing framings rather than pick one.
  • Value-based → make tradeoffs explicit and escalate to a human.
  • Deadlock → emit a structured disagreement report instead of forcing a synthesis.

Failure handling (design requirement, not an afterthought)

• One dead member never sinks the council: drop it and continue with the survivors.
• Chairman failure falls back to the best-ranked member answer.
• Per-member timeouts on every stage.
• State partitioning for parallel writes so one member's write can't corrupt another's.

Deeper failure taxonomy (cascading failures, poison-pill inputs, etc.) is llm-council-failure-modes — this skill only owns "the pipeline keeps running when something breaks."

Frameworks reality (as of mid-2026)

No major framework ships a council primitive — hand-roll it:

• LangGraph: a fan-out/fan-in subgraph plus a synthesis node, for production control flow.
• CrewAI: a crew with an explicit chairman role, for prototyping.
• AutoGen is in maintenance mode; Microsoft directs new work to Microsoft Agent Framework (1.0 GA reported April 3, 2026 — verify against devblogs.microsoft.com/agent-framework before relying on the date).
• Headless coding-CLI councils (e.g. multiple CLI agent instances instead of API calls) are llm-council-harness, not this skill.

Output spec

When this skill is used, produce:

1. The chosen topology (fan-out / sequential / debate) with a one-line justification tied to the task shape.
2. A stage-by-stage diagram or list: what each stage receives, what it returns, and in what format (structured JSON, not prose).
3. The stopping rule (round cap and/or statistical criterion) if debate is involved.
4. The failure-handling plan for a dead member and a dead chairman.
5. Explicit call-outs of anything owned by a sibling skill instead of solving it inline.

Siblings

• llm-council-when — whether a council is warranted at all.
• llm-council-members — which models to select for the panel.
• llm-council-aggregation — the ranking/scoring math.
• llm-council-prompts — exact prompt wording per stage.
• llm-council-cost — token/latency budget tuning.
• llm-council-failure-modes — deeper failure taxonomy.
• llm-council-harness — headless coding-CLI councils.

See references/topologies.md for full citations and the deeper topology/stopping-rule argument.