Contrast a council pattern with a debate pattern as multi-agent coordination shapes
Council fans out independent agents in parallel and aggregates; debate runs them sequentially with each one reading and attacking the previous arguments.
Imagine you cannot decide what to cook for dinner. A council is asking three friends in separate text threads what they would make, then picking the most popular vote; they never hear each other, so the answers are independent. A debate is putting the same three friends in one room and letting them argue, with each new friend reacting to the last one's pitch. The argument can change minds and surface flaws nobody saw alone, but it also takes longer and someone usually has to step in and call the winner. Both can end with a referee picking the final answer, but the way the inputs are gathered is completely different.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Concept explanation~2 min read
Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.
Council and debate are the two canonical ways to put more than one agent on a single question. They look superficially similar (multiple model calls, an aggregator) and they get used as if they were interchangeable cost vs quality knobs. They are not. The structural difference between them is whether each agent sees the others' work before producing its own, and that single bit decides the cost model, the failure modes, and the kinds of problems each pattern actually helps with.
Getting this right matters because both patterns show up in 2026 production code and both fail loudly when applied to the wrong task. A council of three GPT-class agents on a question that triggers a shared training-data bias produces a 3-0 wrong vote. A debate on a noisy arithmetic question produces an elaborate argument and the same noise. Picking the pattern by failure mode rather than by budget is the senior-engineer move.
Mental model: council is parallel monologues; debate is sequential dialogue.
Council: parallel and independent
The mechanism
A council sends the same prompt to N agents in parallel. The agents do not know about each other. Each one produces its own answer, sampled at its own temperature or with its own prompt variant. An aggregator collapses the N answers into one: majority vote for discrete choices, mean or median for numeric, a judge agent that reads all N for free-form prose.
The cost model
- Tokens: roughly
N x answer_size + aggregator_cost. - Latency: one round-trip because the calls run in parallel.
- Context: stays small; each agent gets exactly the original prompt.
This is why councils are cheap relative to debates and why people reach for them first.
What they fix and what they do not
Councils excel at variance reduction. If the failure mode is sampling noise (the same model sometimes gets the answer, sometimes does not), N samples plus aggregation crush that noise. This is the self-consistency idea from Wang et al., and it generalises cleanly to discrete answers and multi-step reasoning.
Councils fail at shared bias. If every agent in the council has the same prior misconception (because they share training data, prompt phrasing, or model architecture), they will all confidently produce the same wrong answer and the aggregator will dutifully promote it. A council of one model sampled at three temperatures is especially prone to this; cross-model councils dilute it but do not eliminate it.
Situations where this technique stops working.
2–4 min · Everything important, quickly.
Real products, models, and research that use this idea.
- Wang et al.'s self-consistency paper is the canonical council pattern: sample N reasoning paths, take majority vote on the final answer.
- Du et al.'s 'Improving Factuality via Multi-Agent Debate' showed debate outperforming single-shot on factual benchmarks at higher cost.
What an interviewer would ask next. Try answering before peeking at the approach.
QHow would you stop a 3-agent debate from converging on the most assertive agent's position instead of the correct one?
Mix model identities so the assertion style varies, force agents to cite evidence for any claim, add a neutral judge with a stricter prompt that down-weights confident assertions without evidence, and run two debates from swapped opening positions and intersect the surviving claims.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Red flags & common mistakes
The phrases that signal junior thinking. Click to expand.
Treating council and debate as interchangeable cost vs quality knobs. Council fails when all agents share a blind spot; debate fails when the strongest arguer wins regardless of correctness. Pick the pattern based on the failure mode, not the budget.
60 second bullets to scan on the way to the call.
Do agents see each other before producing output in council versus debate?
How does the council cost model behave across calls, context, and rounds?
Primary sources. Browse if you want the original framing.
Same topic, related formats. Practice these next.