The Atlas
The 10 layers & 50 sublayers of the intelligence supply chain
The complete atlas of the Supply Chain of Intelligence™ (SCOI) by Anand Arivukkarasu. Every sublayer has a deep dive covering nuances, the startup lens, vertical applications and how to defend it. Stars (★) mark the structurally defensible sublayers of the Defensible Triangle. Canonical source: supplychainofai.com.
L8
Memory
Session, profile, network, institutional, world models.
L8A
Session & Short-Term Memory
What the agent remembers within a conversation, task or run.
L8B
User & Entity Profiles
Persistent, user-controlled memory about people, accounts and entities the agent works with.
L8C ★
Aggregated Network Learning
Cross-customer, privacy-preserving learning — the classic 'gets better as more people use it'.
L8D ★
Institutional Knowledge
The tacit + explicit knowledge of an organization, encoded so agents can use it.
L8E ★
Learned World Models
Long-horizon, environment-level models — how a business, market, road or factory actually works.
L7
Surface
Conversational, visual, embodied, transactional, ambient.
L7A
Conversational
Chat as the interface — the surface AI made famous.
L7B
Visual Interfaces & Media
Canvases, generative UIs, image + video + design surfaces.
L7C ★
Embedded & Embodied AI
AI inside cars, appliances, wearables, glasses, robots — leaves the screen.
L7D
Transaction Surface
Where AI stops being a chatbot and starts moving money, inventory or physical work.
L7E
Async & Ambient Surfaces
AI that runs in the background — inbox, calendar, home, car, city — without a chat window.
L6
Orchestration
Loops, HITL, routing, state, assurance.
L6A
Agent Loops
Plan-execute-reflect loops; ReAct; multi-agent choreographies.
L6B ★
Human-in-the-Loop
Where and how humans check, correct, or approve agent work — and how those signals train the system.
L6C
Role Routing & Task Decomposition
Breaking work into subtasks and routing each to the right skill, tool or sub-agent.
L6D
Context & State Management
What the agent knows at any moment: session, thread, run, task, world.
L6E
Runtime Assurance & Learning Loops
Live monitoring, drift detection and the loops that improve the system from production data.
L5
Execution
Where intelligence does work — tools, playbooks, RAG, actuation.
L5A ★
Domain Execution & Tool Use
The domain-specific tools an agent can actually call — and the semantics of each tool.
L5B ★
Decision Frameworks & Reasoning Scaffolds
The prompted or code-enforced structure that turns raw model reasoning into a decision.
L5C
Retrieval-Augmented Workflows
RAG in production — not the demo — including chunking, graph, freshness and provenance.
L5D ★
Operating Playbooks
Codified 'how work gets done here' — for agents to execute, humans to review, and both to improve.
L5E
Interaction Skills & Actuation
The 'hands' of the agent — browser use, keyboard-mouse, robotic actuation, physical control.
L4
Access
APIs, agent protocols, governance, identity, real-time infra.
L4A
API & Integration Layer
REST/GraphQL/gRPC surfaces, webhooks, connectors and integration marketplaces.
L4B ★
Agent Interface Protocols
MCP, Agent2Agent, tool-use conventions — how agents talk to systems.
L4C
Access Governance & Agent Commerce
Auth, entitlements, rate limits and — increasingly — machine-to-machine payments for agent actions.
L4D
Real-Time Interaction Infrastructure
WebRTC, streaming APIs, low-latency inference plumbing for voice + video agents.
L4E ★
Agent Identity & Provenance
Who / what an agent is, on whose behalf it acts, and the audit trail of its actions.
L3
Gates
Compliance, quality, safety, editorial, distribution.
L3A
Compliance & Export Controls
Export licenses, sanctions, national-security controls on models and compute.
L3B
Quality Gates
Evals, benchmarks, regression suites — the tests that decide what ships.
L3C
Safety, Security & Provenance
Guardrails, jailbreak defenses, PII redaction, content provenance and watermarking.
L3D ★
Editorial Gates
Human editorial control over what models say — style, tone, factual thresholds, brand voice.
L3E ★
Distribution Gates
App stores, model marketplaces, enterprise procurement — who is allowed to reach the user.
L2
Models
Foundation, specialized, retrieval, routing, reasoning.
L2A
Foundation & Multimodal Models
Frontier models from OpenAI, Anthropic, Google, Meta, xAI, Mistral, DeepSeek and their open-weight peers.
L2B
Specialized & Fine-Tuned Models
Domain-tuned models — code, biology, legal, finance, radiology, robotics.
L2C
Embedding & Retrieval
Text, image and code embeddings; rerankers; hybrid search.
L2D
Model Routing & Composition
Cost-quality routers that pick the right model per request; mixture-of-experts style composition.
L2E
Reasoning & World Models
Chain-of-thought, tree search, o-series style reasoning, and true world models for physical AI.
L1
Data
The raw material of intelligence.
L1A
Public & Open Data
Common Crawl, Wikipedia, arXiv, GitHub, open government data.
L1B ★
Proprietary Data
The single most defensible sublayer in the entire stack — data no one else can get.
L1C ★
Behavioral & Sensor Data
Clicks, cursor traces, video, audio, IMU, LIDAR, industrial telemetry.
L1D ★
Outcome Data
What actually happened after the model acted — the ground truth for RLHF, evals and post-training.
L1E
Synthetic & Simulation Data
Model-generated data, simulator rollouts, world-model traces — the fastest-growing L1 sub.
L0
Infrastructure
Silicon, data centers, interconnect, edge compute.
L0A
Silicon & Memory
Accelerators (Nvidia, AMD, TPU, Trainium, custom), HBM, and the memory hierarchy that dominates training cost.
L0B
Data Centers
Powered shells, colo, hyperscale — the physical containers of intelligence.
L0C
Interconnect Fabric
NVLink, InfiniBand, Ethernet-based scale-out — the pipes between accelerators.
L0D
Compute & State Infrastructure
Vector DBs, object stores, feature stores, KV caches — persistent state around ephemeral compute.
L0E
Edge & On-Device Compute
NPUs in phones, laptops, cars, wearables and industrial devices.
L-1
Resources
Energy, water, minerals, people — the physical foundation.
L-1A
Energy & Grid Interconnect
The electricity, PPAs and grid queue positions that gate every AI build-out.
L-1B
Thermal & Water Management
Cooling water, closed-loop chillers and heat rejection that cap real data-center density.
L-1C
Fabrication & Foundry
Leading-edge nodes at TSMC, Samsung and (eventually) Intel — the physical origin of every AI accelerator.
L-1D
Critical Materials & Supply Chain
Gallium, germanium, high-purity quartz, rare earths and the logistics that move them.
L-1E
Skilled Trades & Human Capital
Electricians, HVAC technicians, ML researchers, safety engineers — the humans without whom the stack stalls.