The AI Agent Mastery Iceberg is a tiered map of 30 concepts in agentic AI work, arranged from surface-level basics to advanced practice. Its five tiers, Noob, Beginner, Pro, Expert, and Vibe God, progress from prompting and context windows, through harness tooling such as MCP servers, skills, and hooks, to multi-agent orchestration, self-improving loops, and agent security. Each entry below gives a one-line definition and links to a reference memo on the concept.

Noob

📝Prompting - Writing clear instructions that tell the model what you want
📝Context Window - The max text a model can hold and reason over at once
📝Tokens - The units of text models read, write, and bill by
📝Projects - Workspaces that bundle files and instructions for a shared purpose
📝API Keys - Credentials that let your code call a model directly
📝CLI - Running agents from the terminal instead of a chat window

Beginner

📝MCP Servers - Plug-ins that give agents tools and data via Model Context Protocol
📝CLAUDE.md - A repo file of standing instructions loaded into every session
📝Slash Commands - Saved prompts and actions triggered with a /command
📝Model Routing - Sending each task to the cheapest model that can handle it
📝Subagents - Helper agents spawned with their own context for focused subtasks
📝Prompt Caching - Reusing a stable prompt prefix to cut cost and latency

Pro

📝Git Worktrees - Separate checkouts so parallel agents don't collide on one branch
📝Skills - Packaged instructions an agent loads on demand for a specific job
📝Hooks - Scripts that fire on agent events to enforce rules or automate steps
📝Harness Engineering - Designing the tools, prompts, and loops around a model
📝Background Agents - Agents that keep working while you do something else
📝Headless Runs - Running agents non-interactively from scripts, cron, or CI

Expert

📝Parallel Agents - Running many agents at once on independent pieces of work
📝Progressive Disclosure - Loading only pointers first, deeper context only when needed
📝Eval-Driven Loops - Iterating against tests or scores until output meets the bar
📝Orchestrator-Worker - One lead agent plans and delegates to worker agents
📝Ultrathink - Asking for maximum reasoning budget on hard problems
📝Memory Consolidation - Distilling session learnings into durable, compact memory

Vibe God

📝Self-Improving Loops - Agents that refine their own prompts, skills, and tools over time
📝Tool Poisoning - Malicious tool descriptions or outputs that hijack an agent
📝Self-Healing Tests - Agents that detect and repair broken tests automatically
📝Agent CI/CD - Pipelines where agents build, review, test, and ship code
📝Agents Managing Agents - Supervisory agents that spawn, monitor, and steer other agents
📝Multi-Model Orchestration - Coordinating different models, each on what it does best

Contexts

📝Claude Code Glossary