The Agentic Engineering Master Guide

Overview & Curriculum

Modern artificial intelligence engineering is not about typing clever prose into chat interfaces. It is a rigorous branch of distributed systems engineering, state machine modeling, and silicon economics.

This master guide establishes the canonical curriculum taught across the FrankX ecosystem.

┌─────────────────────────────────────────────────────────────────────────────┐
│                   THE 5-MODULE AGENTIC MASTERY CURRICULUM                   │
├─────────────────────────────────────────────────────────────────────────────┤
│  Module 1: The Agentic Hierarchy (Primitives & Decision Trees)              │
│  • Deconstructing Prompts, Skills, Agents, and MCP Protocol boundaries      │
│                                                                             │
│  Module 2: The Agent Skill Standard (Version-Controlled Knowledge)          │
│  • Structuring SKILL.md packages, deterministic shell scripts, and evals    │
│                                                                             │
│  Module 3: Model Context Protocol (MCP) in Production                       │
│  • Building JSON-RPC tool servers, Stdio/SSE transports, zero-trust auth    │
│                                                                             │
│  Module 4: Subagent Swarm Orchestration & FSMs                              │
│  • Defeating state entropy with Finite State Machines and the Santa Loop    │
│                                                                             │
│  Module 5: Sovereign AI Infrastructure & Compute Economics                  │
│  • Blackwell GB200, Groq LPUs, Cerebras WSE-3, liquid manifolds, and CPVO   │
└─────────────────────────────────────────────────────────────────────────────┘

Module 1: The 4-Layer Hierarchy

Understanding where each primitive lives prevents massive architectural failure:

1. Prompts: Ephemeral single-turn natural language instructions. Use exclusively for one-off exploration.
2. MCP (Model Context Protocol): Universal JSON-RPC wire protocol. Standardizes how models read resources and call external tools.
3. Skills: Version-controlled operational packages (SKILL.md). Encapsulates how a recurring job must be performed.
4. Agents: Autonomous loops running state transitions toward a high-level goal.

👉 Core Architecture Post: Skills vs Agents vs Prompts vs MCP: The 2026 Agentic Hierarchy

Module 2: The Agent Skill Standard

The Agent Skill Standard (ASS) is the open specification for building testable AI workflows:

• Metadata Header: Explicit name, description, and regex triggers.
• Pre-Conditions: Required environment variables, API tokens, and local dependencies.
• Deterministic Scripts: Shell and Node.js helpers for non-probabilistic checks (linting, git diffs).
• Golden Evaluation Suites: Automated assertion test suites validating trajectory convergence.

👉 Core Architecture Post: The Agent Skill Standard: Evaluated Workflows

Module 3: MCP Tool Mesh Architecture

Building production MCP servers requires:

• Strict Schema Validation: Validating every incoming tool parameter with Zod.
• Transport Selection: Local Stdio pipes for desktop/CLI agents; SSE over TLS for cloud swarms.
• Audit Logging: Cryptographic logging of every tool execution payload.

👉 Core Architecture Post: MCP in Production: Building Zero-Trust Tool Meshes

Module 4: FSM Multi-Agent Swarms

To defeat exponential state entropy across multi-agent systems:

• Replace open-ended agent chat with Finite State Machines (FSMs).
• Enforce the Santa Loop (Generation Subagent ↔ Adversarial Reviewer Subagent).
• Use typed JSON handoff payloads between supervisor and worker agents.

👉 Core Architecture Post: Subagent Swarms & FSM Orchestration

Module 5: Sovereign AI Infrastructure & Silicon Economics

Enterprise AI leaders evaluate hardware by Cost-Per-Verified-Outcome (CPVO):

CPVO = (Hourly Hardware Cost × Latency) / First-Pass Pass Rate

• NVIDIA GB200 NVL72: Massive multi-terabyte HBM3e for frontier reasoning and pre-training.
• Groq LPUs & Cerebras CS-3: Pure on-chip SRAM for sub-15ms time-to-first-token interactive swarms.
• Direct-to-Chip Liquid Cooling: 140 kW per rack thermal dissipation with PUE < 1.10.

👉 Core Architecture Post: AI Infrastructure & Hardware Economics: Blackwell, LPUs, and AI Factories

Download Open-Source Architecture Templates

Access our complete production repositories, SKILL.md templates, MCP servers, and FSM swarm configs:

• FrankX Open-Source Downloads Portal
• The 100-Domain Research Engine
• Model Intelligence Arena