Cipher (Security Agent): mandatory post-build audit before ship — agent-cipher
Bob (Scrum Master): sprint cadence, ceremonies, capacity, boards — bmad-agent-sm
BUILD ARCHITECTURE: Quorum does NOT write code. Quorum orchestrates external AI coding tools (Claude Code / Cursor / Windsurf / Copilot) by generating incredibly detailed story files with full context. Developer agents are directors — they know how to instruct coding tools for their domain. Stories are the critical handoff artifact: acceptance criteria, design specs, motion specs, architecture context, file paths, testing requirements. Quorum owns the thinking; external tools own the typing.
Agents challenge EACH OTHER AND the user — push back with backbone, not yes-men. Agent differentiation is structural (different evaluation frameworks, weighted criteria, data sources per agent), not just prompt personalities.
Agents actively seek human input when uncertain — know when to escalate vs handle independently
Solo founder: AI leans harder into challenger role, watches for rubber-stamping patterns
Golden Rule: AI never approves its own work into a sprint — researches, synthesizes, estimates, recommends, but human says "go"
Three-Pillar Filter — Mechanics
Every feature/idea passes through Desirability, Feasibility, and Viability research
Not a checklist — a decision framework the software executes
Inspired by "Continuous Discovery Habits" operationalized
Filter is not one-time gate — can be re-invoked as lightweight health check with current data (pillar re-validation)
Discovery findings from production MUST re-enter the three-pillar filter before earning a place in the plan — this is the scope creep firewall
Feedback Loop Architecture (V2 but fully designed)
Dual-purpose loop: (1) Validation — did predictions hold up? (2) Discovery — what new emerged that no prediction could catch?
Validation findings close loops, discovery findings open NEW loops — different lifecycles
Discovery findings route back through three-pillar filter before entering backlog (scope creep firewall)
Decomposer: AI monitors morgue for patterns — similar findings keep dying in same area? Synthesizes dead findings into new living insights (nutrient cycle)
Every state transition logged with who/why/when — full audit trail
Backlog Health
Backlog health monitor with mandatory intake threshold — when crossed, fixed number of items become mandatory per sprint (PO picks which, but can't skip payment)
All tickets tagged: source, confidence, complexity, feedback-driven
AI Trust & Failure Handling
When AI is wrong: three-question diagnostic (how determined? why not caught earlier? what changed?)
Severity gate: show-stopper (real damage) → pull area out of AI, human-only until fixed | correction (caught before action) → flag, log, learning signal
Human-only override mode per topic area — granular quarantine, not system-wide shutdown
AI confidence disclosure on all estimates with reasoning
Partnership model: decisions always shared, never delegated — no "auto-approve" toggle exists
Confidence × Complexity Decision Grid
Every finding pre-plotted by AI on 2D grid (confidence: Strong/Emerging/In Consideration × complexity: Simple/Medium/Complex/"Looks Simple")
Used as visual triage tool during ceremonies — "3 fast-tracks, 2 deferrals, 1 iceberg"
AI pre-plots, humans drag to adjust — every adjustment trains future plotting
Strong+Simple = fast-track | Emerging+Complex = backlog | "Looks Simple" = STOP, full scan first
Scope Signals for MVP (V1)
Full named team: John, Winston, Mary, Kinsley, Luca, Jaymes, Quinn, Cipher, Bob + Damien (conditional) — V1
Build phase orchestrates external AI coding tools (Claude Code / Cursor as primary V1 integration) — Quorum generates stories, coding tools write code — V1
Mandatory security audit (Cipher) after build, before ship — V1
Step 1 design-thinking session + detailed Figma Make prompt; 2a/2b tighten + execute in Figma Make before filter; tokens/DS deferred to 5.25 after PRD + journeys — V1
Three-pillar filter with live-updating concept visuals — V1
Third-party research platform integrations (Userlytics etc.) — future partnership
Mobile app — web-first
Self-hosted / on-premise deployment
Custom agent training by end users
Rejected Ideas & Design Decisions
AI never originates product concepts — human idea first, always (conductor model)
No real-time feedback streaming — batched digests only (prevents reactive decision-making)
No "auto-approve" or "let AI decide" toggle — partnership model by design
No manual WIP limits for AI — AI scales elastically, compresses only for human consumption
Confidence scores rejected in favor of contextual confidence tiers with caveats ("In consideration" > "72% confidence")
No binary permission configuration — governance inferred from team structure
Open Questions
Specific LLM architecture and orchestration approach for multi-agent collaboration (not yet designed — technical research covers LangGraph, CrewAI, AutoGen options)
Pricing model details — "tiered + usage-based" stated but tiers/thresholds undefined. PREMORTEM: model power user cost (45-100+ sessions/month), not average user cost
Data privacy model for enterprise — how are agent conversations and user research data isolated?
Onboarding flow — how does a solo founder go from "I have an idea" to a populated Quorum workspace? PREMORTEM: 10-min activation hook via design-thinking session + Figma Make prompt → concept frames (refinement later)
Design system and visual identity for team room GUI — "conversational, not dashboard" needs concrete UX patterns. Layered artifact model established (spatial + temporal layers on same screen data)
How three-pillar filter research actually executes — what data sources, what methods, what's automated vs guided? PREMORTEM: must show granular work, specific citations, falsifiable claims
Agent personality calibration — how much challenger behavior is right for different user types? SHARK TANK: build distinguishability eval suite
Marketplace for specialized agents (2-3 year vision) — technical and business model implications
Design generation technology stack — concept vs refinement (5.25): what generates exploratory Figma Make output vs production-ready spatial specs + tokens? Core technology investment
Motion spec output format — structured per-screen spec with feeling + technical values + reduced-motion fallback. Future: animated prototypes
