AI AGENT DESIGN · NEW TRACK

Design agency before autonomy.

An agent is more than an LLM with tools. Learn to decide when agency is warranted, define the boundaries of action, allocate authority, design closed-loop behavior, and create the evidence needed for deployment.

Start Lesson 01 View track

agent_design / authority_model

DESIGN QUESTION

What decisions can this system make without asking a human?

OBSERVEshipment + weather + SLA

REASONallowed

RECOMMENDallowed

EXECUTEbounded + reversible only

COMMIThuman approval required

TRACK THESIS

Agency is a system design choice.

The useful question is not “Can the model do this?” It is “What observation, decision, action, verification, and stopping authority should the system receive?”

Observe

What information can the system retrieve, and what evidence must be present before it acts?

Decide

Which choices may be delegated, and which decisions stay with a named human owner?

Act + verify

What tools can change the world, how are side effects bounded, and how does the system verify the result?

AI AGENT DESIGN · FOUNDATION

Seven lessons before framework code.

The first sequence creates a reusable design record before the learner chooses LangGraph, an agent SDK, or any other orchestration layer.

LESSON 01

Agent or Workflow?

Distinguish deterministic automation, an LLM workflow, and an agent from the structure of the decision—not from the marketing label.

OUTPUTAgent Suitability Record

AVAILABLE LESSON 02

Define the Agent Boundary

Set goal, knowledge, tool, authority, cost, time, and failure boundaries.

OUTPUTAgent Boundary

AVAILABLE LESSON 03

The Agency Budget

Allocate permission to read, analyze, recommend, modify, commit, communicate, spend, and delete.

OUTPUTAuthority Matrix

AVAILABLE LESSON 04

Design the Agent Loop

Connect observation, interpretation, choice, action, verification, escalation, and stopping.

OUTPUTAgent Loop Map

AVAILABLE LESSON 05

Tool Design

Turn APIs into bounded capabilities with typed inputs, authorization, idempotency, and explicit failure behavior.

OUTPUTTool Contract

AVAILABLE LESSON 06

State and Memory

Separate working state, episodic memory, semantic knowledge, and authoritative external records.

OUTPUTState & Memory Plan

AVAILABLE LESSON 07

Context Engineering

Assemble the smallest, freshest, most decision-relevant context for each step of the agent loop.

OUTPUTContext Assembly Policy

AVAILABLE

START HERE

Lesson 01 · Agent or Workflow?

Classify three real operating patterns, score an agent candidate, and leave with an Agent Suitability Record.

Begin →