Quorum
Multi-agent intelligence for any question. SMEs debate, challenge each other, and converge on what survives scrutiny. One command.
/quorum "your question here"
Built by qinnovate | Full docs on GitHub
Four Tiers
| Command | What Happens | Agents |
|---|---|---|
/quorum "question" | 5 SMEs debate, supervisor synthesizes | 5 (research-backed default) |
/quorum "question" --max | Full adversarial-driven convergence, teams if needed | 7-15 recommended (supervisor decides, user can override higher) |
/quorum "question" --reviewers | Top-down sequential review pipeline, auto-decide, surface taste calls only | 3-5 phases (topic-driven) |
/quorum "question" --set 200 | Custom scale — swarm auto-engages at 20+ | User-defined |
That's it. The supervisor handles everything else: mode, structure, rigor, research, teams.
Why These Numbers
| Tier | Agents | Research Basis |
|---|---|---|
| Default | 5 | Woolley et al. 2010 (collective intelligence peaks with equal conversational turns in small groups); Du et al. 2023 (3-agent AI debate optimum + supervisor + adversarial = 5) |
| Max | 7-15 (recommended) | 7 = synchronous ceiling before conversational inequality (Dunbar layer 1); 15 = Delphi panel optimum (Linstone & Turoff 2002). User can request more — supervisor scales to --max N if specified |
| Set N | User-defined | At 20+, swarm architecture auto-engages: MECE taxonomy partitioning, environment-based coordination, pattern detection |
Mandatory Adversarial Minimum: 2
Every panel of 5+ agents includes at least 2 adversarial agents. Not 1.
- Asch (1951): A single dissenter reduces conformity from 32% to 5%
- Moscovici (1969): A minority of 2 establishes a credible pattern — 1 is dismissed as eccentric
- Nemeth (2001): Assigned devil's advocacy makes people MORE entrenched. Critics must hold authentic positions with counter-proposals
- Schweiger (1986): Critics who propose counter-plans produce 34% higher decision quality than critics who only attack
Only 8 Optional Flags
| Flag | Why It Can't Be Auto-Detected |
|---|---|
--artifact PATH | Supervisor can't know which file you mean |
--ratify | User wants human-in-the-loop approval before verdict is final |
--reviewers | User wants vertical sequential review, not horizontal debate |
--no-web | Privacy choice — stops web searches. Note: does not block Codex CLI prompt improvement (v7.3.0) when vagueness gate fires |
--ponder | User explicitly wants Q&A before the swarm runs |
--dry-run | User wants to see the plan without spending tokens |
--diverse | Use multi-model panel (Gemini + Codex alongside Claude). Auto-enabled with --max when CLIs detected |
--no-diverse | Suppress multi-model auto-detection under --max |
Everything else is auto-detected by the supervisor:
| What You Say | What Fires | How It's Detected |
|---|---|---|
| "Should we use X or Y?" | Dialectic (2 agents, Socratic rounds) | Binary question pattern |
| "Build a REST API for..." | Hackathon (PRD + TDD + Ralph loop) | Implementation intent: "build", "implement", "create", "add feature" |
"Review this" + --artifact | Review mode (agents analyze the file) | Artifact present + review/audit/validate language |
| "What am I missing about..." | Explore mode (reframe the question) | Meta-question / exploratory language |
| "EEG auth methods landscape" | Research mode (web search + synthesis) | Open knowledge question without artifact |
Any question at --max | Adversarial-driven convergence (converse internally) | --max always uses iterative rounds |
Any question at --set 20+ | Swarm (MECE taxonomy + environment) | Agent count ≥ 20 |
| 3+ domains detected | Teams (internal deliberation, cross-challenge) | Supervisor detects domain count in Phase 0.5 |
Forecasting question at --set | Prediction mode (sentiment + coalitions) | "Will X happen", "by 2028", future-tense patterns |
Examples
# Quick opinion — 5 agents, done in 2 minutes
/quorum "Should we use PostgreSQL or DynamoDB for our new service?"
# Stress-test a decision — full adversarial-driven convergence
/quorum "Should we build or buy our auth system?" --max
# Build something — auto-detects hackathon mode, generates battle-tested PRD
/quorum "Build a REST API for user auth with JWT" --max
# → Supervisor detects "Build" → generates PRD with TDD + acceptance criteria
# → Converse stress-tests the PRD (Architect, Breaker, TDD Enforcer, Pragmatist, Judge)
# → Outputs: _swarm/prd-user-auth.md (battle-tested PRD)
# Review a document
/quorum "Review this proposal for risks" --artifact proposal.md
# Deep philosophical question — auto-routes to dialectic
/quorum "Should we open-source our core product?"
# Research landscape — auto-routes to web research + synthesis
/quorum "Complete landscape of EEG-based authentication methods"
# Massive scale prediction — swarm auto-engages
/quorum "Will BCI startups consolidate or fragment by 2028?" --set 200
# Private, no web searches
/quorum "Evaluate our internal security posture" --artifact audit.md --no-web
# See the plan before spending tokens
/quorum "Microservices or monolith?" --max --dry-run
Hackathon Mode (Auto-Detected)
When you say "build", "implement", "create", "scaffold", "write a", "set up", or "add feature", the supervisor auto-triggers the hackathon pipeline. No flag needed.
/quorum "Build a REST API for user auth with JWT" --max
What happens:
- Supervisor detects implementation intent ("Build") → triggers hackathon pipeline
- Decomposition agent generates a PRD with TDD enforcement:
- Exact file paths for every file created or modified
- Bite-sized tasks (one action each, 2-5 minutes)
- Each task: write failing test → verify fail → implement → verify pass → commit
- Machine-verifiable acceptance criteria (not "works correctly" but "returns 200 with valid JWT containing user_id claim")
- Adversarial-driven convergence stress-tests the PRD (only with
--max):- Architect: "Are the boundaries right? Missing abstractions?"
- Breaker: "Which acceptance criteria are ambiguous? Edge cases?"
- TDD Enforcer: "Is every task actually testable? Assertions specific enough?"
- Pragmatist: "Is this over-engineered? Can tasks be eliminated?"
- Judge: Computes convergence score. Declares READY or sends back for revision.
- Output:
_swarm/prd-{name}.md— battle-tested PRD ready for implementation
The Ralph loop executes each PRD task with fresh context:
- Reads PRD + progress.md + AGENTS.md
- Picks highest-priority incomplete task
- TDD: test → fail → implement → pass → commit
- Updates progress.md with learnings
- Every 3 tasks: Quorum review catches regressions, skipped tests, architecture drift
- Repeats until all tasks checked off
Note: Ralph loop is local-only. Not included in the published marketplace version.
With vs without --max:
/quorum "Build X" | /quorum "Build X" --max | |
|---|---|---|
| PRD generation | 5 agents (lighter) | 7-15 agents (full decomposition) |
| Stress-test | No adversarial review | Full convergence (Architect + Breaker + TDD Enforcer + Pragmatist + Judge) |
| Output quality | Good for small features | Battle-tested for production systems |
Trigger keywords: build, implement, create, add feature, scaffold, write a, set up — anything that signals "I want code output, not analysis."
No --hackathon flag exists. Same capability, zero cognitive load.
Vagueness Gate (Auto-Triggered)
If your prompt is too vague to produce a useful verdict, the supervisor stops and asks before spending tokens. This fires automatically — no flag needed.
What triggers it (any 2+):
- No