FIELD GUIDE · v1.0

9 Building Blocks of Agentic AI

The vocabulary every engineering team needs before shipping agents to production. Nine visual vocabulary cards: Agentic Loops, MCP, Subagents, AI Gateway, Inference Economics, Evals, Guardrails, Observability, and The Bitter Lesson.

01
Pattern · Control Flow

Agentic Loops

Plan → act → observe → reflect — repeated until a goal is met. The base unit of agentic behaviour.

Plan
Act
Observe
Reflect
02
Protocol · Tool Integration

MCP — Model Context Protocol

One interface, every tool. Standardised connectors instead of N×M custom integrations.

03
Architecture · Multi-Agent

Subagents & Multi-Agent Systems

Delegated reasoning, isolated context. An orchestrator splits work across specialised subagents and merges the result.

04
Infrastructure · Control Plane

AI Gateway

One control plane, many models. Auth, rate-limits, logging, fallback — centralised in front of every provider.

05
Economics · Cost Model

Inference Economics

Tokens are the unit of cost. Cache-hits land at ~10 % of the price; misses pay full freight.

06
Quality · Measurement

Evals

What you can’t measure, you can’t ship. Test cases drive model output through a judge — pass/fail aggregates into shipping signals.

07
Safety · Brand Protection

Guardrails

Input filter blocks PII and prompt injections; output filter blocks jailbreaks, leaks, and policy violations.

08
Operations · Visibility

Observability

You can’t fix what you can’t see. Traces, logs, and metrics from every agent step land on one dashboard.

09
Principle · Sutton, 2019

The Bitter Lesson

Handcrafted rules are brittle and lose long-term. Methods that scale with compute & data win — every time the cycle repeats.