The Eight Deliberation Modes, Explained

Consilium ships eight deliberation modes. They're not marketing skins on the same logic — each one is a distinct state machine in the deliberation graph with a different round count, transition shape, and judge behavior. Picking the right one matters. Picking the wrong one wastes spend or produces a thinner answer than the topic deserves. Here's what each mode actually does, when to reach for it, and what it costs.

The lineup

Round counts are the actual values in MAX_ROUNDS_BY_MODE in apps/agents/src/features/deliberation/deliberation_graph.py.

quick — single round, no debate

Each model produces one independent answer; the judge picks the best one. There is no cross-examination round. This is the cheapest mode and the only mode where you are essentially getting a ranked best-of-N from your council. Use it when you want multiple model perspectives but the question is simple enough that none of them is going to challenge another's answer productively.

council — the default

Three rounds: independent analysis, cross-examination, rebuttal-and-refinement. The judge runs the 5-phase synthesis on the round-3 outputs. This is what we recommend for most non-trivial reasoning tasks: long enough to surface disagreement, short enough not to drag.

deep — five rounds with sub-agent research

Adds two more rounds beyond council, plus optional sub-agent research for context-heavy questions. Worth the extra spend when the topic is genuinely contested and the room is split after round 3. Convergence detection still applies, so a deep-mode debate that locks in early ends early — you only pay for the rounds you needed.

blind — names hidden until scored

Same shape as council, but model identities are stripped before each round so models judging round-1 outputs in round 2 don't know which one came from Claude vs. GPT vs. Gemini. Useful when you suspect anchoring to a particular brand — for example when ranking responses to a prompt where the "safe" answer is the one a more cautious model wrote.

redteam — attack/defend cycle

Designed for security review. The flow is asymmetric: one subset of models attacks (generates exploits / failure modes / adversarial inputs), another defends, and the judge categorizes findings into five severity-ranked dimensions (security, bugs, performance, quality, edge cases). Despite the single-round count, redteam runs longer than council in wall-clock time because the attack and defend phases are sequential and each pulls more tokens than a normal round-1 generation.

jury — ranked-choice voting

Three rounds. After round 3, every model casts a ranked ballot over the candidate answers. We aggregate using Borda count for the headline score and run a Condorcet check for cycle detection. When the room produces a consistent preference order, jury surfaces it; when there's a cycle (A beats B beats C beats A), the judge produces a weighted synthesis that the dissent report explicitly flags. Use jury when there are clearly multiple defensible answers and you want the council's collective ranking instead of a single verdict.

market — prediction-market aggregation

Five rounds with confidence-weighted voting. Each model produces a probability claim and a justification; the judge aggregates the probabilities the way a prediction market would — weighted by each model's calibrated confidence on similar past questions. This is the right mode for forecasting ("will WebAssembly replace Docker for serverless within 3 years?") where the deliverable is a probability with reasoning, not a yes/no.

auto — let the engine pick

Auto runs a small classifier on the topic in front of round 1 and routes to one of the seven explicit modes. The decision is surfaced in the SSE stream as a routing:decidedevent so you can see what it chose and why — auto is not a black box, it's a default with an audit trail. If the classifier is uncertain it falls back to council, which is the safest non-trivial mode.

How to choose

Three rules of thumb that work in practice:

• If the topic has a fact-of-the-matter answer, start with quick. You don't need debate when one round of best-of-N is enough.
• If the topic has tradeoffs, council. The cross-examination round is where tradeoff analysis actually happens.
• If the deliverable is structured (a vote, a probability, a security report), pick the matching specialized mode — jury, market, or redteam respectively.

And if you genuinely don't know, auto is fine. It's what we use as the default for unclassified incoming traffic.