@meta
  v: 1
  route: /docs/how-it-works
  generated: 2026-06-17T00:00:00Z
  ttl: 1d

@intent
  purpose:    Explain Roundtable's two modes — advisory panel vs. chaired deliberation to consensus
  audience:   developer, ai-agent, decision-maker
  capability: understand, choose-mode

@state
  modes[2]{id,rounds,heads_see_each_other,output}:
    advisory,1,no,every independent take plus a synthesized recommendation
    deliberation,N,yes (after round 1),revised answers until consensus plus a synthesized verdict
  chair: a dedicated Claude seat that adjudicates; it does not compete
  chair_verdicts[2]{verdict,meaning}:
    CONSENSUS,heads have converged; council stops early
    CONTINUE,meaningful disagreement remains; run another round
  rounds_flag: --rounds N is an upper bound; consensus ends it sooner
  research_mode: --research lets tool-capable heads gather fresh context; off by default

@actions
  - id: run_advisory
    method: GET
    href: https://roundtable.sh/docs/cli-reference
    command: roundtable ask "What's the biggest risk in this rollout plan?"
  - id: run_deliberation
    method: GET
    href: https://roundtable.sh/docs/cli-reference
    command: roundtable ask "Postgres or SQLite for this workload?" --rounds 3

@context
  > Advisory is for breadth — diverse, uncorrelated takes, fast. Deliberation is for high-stakes
  > calls where you want the models to reconcile their disagreements: each round every head sees
  > the whole table and revises, while a Claude chair decides whether the council has converged.

@nav
  self:      /docs/how-it-works.agent
  parents:   [/]
  peers:     [/docs/get-started.agent, /docs/the-heads.agent, /docs/cli-reference.agent, /docs/mcp-and-harness.agent, /docs/agents.agent]
  drilldown: /docs/the-heads.agent