01 / MODEL CONTEXT PROTOCOL
MCP
Model Context Protocol
Open standard enabling AI agents to connect tools, APIs, and data sources through a unified interface.
Claude connects to GitHub via MCP to read and write code autonomously.
The vocabulary of agentic AI
12 concepts every builder should know.
Agentic systems aren't magic — they're composable patterns. Memory, tools, loops, guardrails. Get the vocabulary right and the architecture follows. A reference for builders, operators, and decision-makers.
12
Core concepts
3
Layers — Cognition · Action · Safety
∞
Combinations
The vocabulary
02 / PERCEIVE → PLAN → ACT → OBSERVE
Agent Loop
Perceive · Plan · Act · Observe
The continuous reasoning cycle an AI agent runs — sensing input, forming a plan, executing actions, then observing results.
An agent reads an error log, plans a fix, runs the code, and checks if tests pass.
03 / AGENT CAPABILITIES
Tool Use
Agent Capabilities
The ability of an AI model to call external functions — APIs, code runners, browsers — to act on the real world.
Agent calls a weather API mid-conversation to answer "Should I fly to NYC tomorrow?"
04 / AGENT MANAGER
Orchestrator
Agent Manager
The top-level agent that breaks goals into subtasks and delegates them to specialized subagents.
A coding orchestrator dispatches a test agent, a lint agent, and a deploy agent in parallel.
05 / SPECIALIZED WORKER
Subagent
Specialized Worker
A focused AI agent that executes one specific task within a larger multi-agent pipeline.
A "summarizer" subagent condenses 50 research papers before the main agent synthesizes insights.
06 / SHORT-TERM & LONG-TERM
Memory
Short-term & Long-term
How agents retain and retrieve information — in-context (temporary) or via external stores (persistent).
Agent recalls your coding style from a previous session stored in a vector database.
07 / REALITY TETHERING
Grounding
Reality Tethering
Connecting AI outputs to verified external data sources to reduce hallucination and improve accuracy.
Agent cites live stock prices from Bloomberg API instead of generating numbers from training data.
08 / SAFETY LAYER
Guardrails
Safety Layer
Rules and constraints that prevent agents from taking harmful, unauthorized, or out-of-scope actions.
Agent is blocked from deleting production databases even if instructed to "clean everything up."
09 / SAFE EXECUTION
Sandboxing
Safe Execution
Isolated environment where agents run code without risk of affecting the host system or production.
Claude Code runs user scripts in a Docker container before applying changes to the actual repo.
10 / APPROVAL GATE
Human-in-the-Loop
Approval Gate
A design pattern where agents pause and request human confirmation before executing high-stakes actions.
Agent drafts a client email but waits for your approval before hitting send.
11 / WORKING MEMORY LIMIT
Context Window
Working Memory Limit
The maximum amount of text an agent can read and reason over in a single interaction — its attention span.
A 200K token window lets an agent read an entire codebase before writing a single line.
12 / COLLABORATIVE INTELLIGENCE
Multi-Agent
Collaborative Intelligence
A system where multiple specialized AI agents collaborate, each handling different tasks to solve complex goals faster and more reliably.
One agent researches, one writes, one fact-checks — all coordinated by an orchestrator.
How it fits together
An orchestrator drives the loop, dispatches subagents, uses tools through MCP, grounds outputs in real data, and pauses for human approval at high-stakes gates.
Goal → Orchestrator → Plan → Tool / MCP → Subagent → Observe → HITL ✓
MEMBER · FREE
Read the full article or download as PDF
The full article and the PDF are member content. Magic-link login, no credit card, no risk — both available immediately.