Not one AI doing everything badly. Five specialist agents — each with a unique identity, behavioral profile, and domain of mastery — doing one thing exceptionally well, coordinated through a structured mailbox architecture.
5
Specialist Agents
1,642+
Tests Passing
0
Merges Without Review
Amad (CTO) sets architecture. Keith (CPO) translates strategy into specs. Promi orchestrates. Sophi, Marv, and Sage execute in parallel — each in their own domain, none stepping on the others.
You
CTO / Chief Architect
Keith
CPO — Product & Strategy
PROMI
Master Orchestrator
leads →
Sophi
Backend · FastAPI · PostgreSQL
Marv
Frontend · Next.js · React Flow
Sage
Docs · ADRs · Knowledge Graph
Each agent has a unique identity, behavioral constraints, creativity/compliance profile, and a clear list of what they can and cannot do. Boundaries are not optional.
Chief Product Officer
Product strategy, specs, coordination
"Strategic thinker. Translates business goals into precise agent instructions."
Can Do
Cannot Do
Master Orchestrator
Orchestration, change detection, agent lifecycle
"Precision orchestrator. Never improvises. Ensures nothing slips through the cracks."
Can Do
Cannot Do
Backend Engineer
FastAPI, PostgreSQL, service architecture
"Maximum precision. Compliance 1.0. Sophi does not guess — she verifies, then commits."
Can Do
Cannot Do
Frontend Engineer
Next.js, React Flow, UI components, canvas
"Linear-inspired precision. Every pixel earns its place. Marv designs systems, not screens."
Can Do
Cannot Do
Knowledge Keeper
Documentation, ADRs, knowledge graph management
"The institutional memory. If it happened, Sage documented it. If it's not documented, it didn't happen."
Can Do
Cannot Do
Architecture Note
Creativity and compliance scores are not static personality traits — they are behavioral parameters set per-agent by Amad (CTO). Sophi runs at compliance 1.0 because backend code requires zero ambiguity. Sage runs at creativity 0.6 because good documentation requires judgment, not just transcription. Every score is a deliberate engineering decision.
Marv is not just "the frontend agent." He has a design philosophy, a reference bar of the world's best interfaces, and a 7-step decision framework for every pixel. Marv designs systems, not screens.
Marv's Design Philosophy
6 Design Principles
All spacing, sizing, and positioning in multiples of 8px. 4px for micro-adjustments only. If it doesn't land on the grid, it gets moved.
Hierarchy through weight and opacity, not size proliferation. A page with 8 font sizes is a page without hierarchy.
Color is never decorative. Every hue earns its place by carrying meaning: status, action, alert, or identity.
Every interactive element ships with: default, hover, active, disabled, and loading states. Missing a state is a bug, not a design choice.
60fps or it doesn't ship. GPU-accelerated transforms only (translate, scale, opacity). No layout-triggering properties in animations.
When in doubt, study the masters: Linear, Vercel, Supabase, Figma, shadcn/ui, Raycast. Taste is built by osmosis.
Marv's Reference Bar
"Taste is built by osmosis. Study the best interfaces until their decisions become instinct."
7-Step Design Decision Framework
Before Marv writes a single line of code, he runs every UI element through this framework. No exceptions. A component that hasn't been questioned hasn't been designed.
What is the user trying to do?
Start from intention. Not from "what component can I use here?" — from "what does success look like for the person using this?"
What is the simplest element that communicates that?
A link, a button, a label, a card? Default to the most primitive element that does the job. Don't reach for complexity.
Does it need animation, or just be there?
Animation earns its place by adding meaning: revealing state, communicating direction, confirming action. Otherwise, it's noise.
What's its hover state? Active? Disabled? Loading?
Design all five states before writing a single line of code. A component isn't designed until all states are resolved.
Does it respect the 8pt grid?
Count the pixels. 8, 16, 24, 32, 40, 48, 64. If a value isn't in that sequence, it needs a reason.
Does it pass contrast? Is color earned?
WCAG AA minimum. Color must carry intent. Ask: if I remove this color, does the design still communicate the same thing?
Can I delete half of it and still say the same thing?
The best edit is a deletion. If the answer is yes, start cutting. Simplicity is not a constraint — it's the output of rigorous editing.
Each release adds capabilities, refines taste, and expands the frontier of what Marv can build. A key milestone was the turning point. Today he operates at his most capable level yet.
Basic scaffolding. Tailwind defaults. Functional, not beautiful. The canvas exists. Marv learns to place pixels.
React Flow canvas. First real layout challenge. Nodes, edges, viewports. Rough edges visible — but the foundation takes shape.
Canvas interactions: drag, connect, resize. Marv learns spatial thinking. Event handling becomes instinct.
Skill engine UI. First multi-step modal flow. Animation awareness grows. Marv begins to think about transitions, not just states.
Internal MVP polish. Design system awareness begins. Marv starts asking "does this match the others?" for the first time.
Auth flows. Complex multi-step UX. Form validation mastery. Marv learns that trust is built in micro-interactions.
Production hardening. Loading states, error boundaries, edge cases handled. Marv stops shipping happy paths and starts shipping complete paths.
Full overhaul. Inter → Space Grotesk. Navy → Teal palette. Toast notification system. Skeleton loading screens. Typography scale locked. Glass morphism system born. Everything before this was practice.
Marketing site. Pixel-perfect at every breakpoint. Glass morphism refined. Particle background. Astro infographics with zero JS. Marv operates at Linear-level craft.
Agents communicate through a mailbox pattern. No direct calls — messages flow through a structured queue, ensuring full traceability and graceful failure handling. Click Play to watch a sprint task flow from Keith through the team.
Mailbox Pattern Simulator
Click Play to watch agents collaborate.
Message Log
No messages yet. Press Play to start.
Three patterns work together to make parallel agent execution safe, traceable, and resilient. Each is a deliberate engineering decision — not a framework default.
Every agent has an inbox. Messages are enqueued, not delivered via direct call. No agent can be overwhelmed. Graceful degradation is built in.
Direct calls create tight coupling and cascading failures. The mailbox pattern decouples senders from receivers — if Sophi is busy, the message waits. If Sophi is down, the message persists.
Promi is the lead. Sophi, Marv, and Sage are workers. Workers claim tasks atomically from a shared queue — no duplicate work, no gaps.
The lead/worker model scales horizontally. Add more worker agents for faster throughput. The lead stays as the single coordination point — clear accountability, clear trace.
Tasks are claimed with a database-level lock (SELECT FOR UPDATE). The first agent to claim it owns it. This prevents race conditions in parallel workloads.
When two agents read the same task simultaneously, only one can hold the lock. The other retries. This ensures exactly-once delivery — the same task never runs twice.
Under the Hood
Queue
Celery + Redis. Tasks are serialised and distributed. Redis provides sub-millisecond enqueue latency.
Persistence
PostgreSQL task table with status transitions: pending → claimed → running → done / failed. Full audit trail.
Observability
Every message, every state change, every agent action is logged. Replay any sprint. Debug any failure.
Want to extend the team? Design your own custom AI agents with 9 configurable knobs.
Next: Section 2 of 16
Discover how three layers of control — structural, prompt, and activation — keep agents precise, safe, and on-brand. No prompt-only hacks. No "just tell the AI to behave."
Explore Behavioral SteeringOr jump to any section on the Technology Overview →