Level 4
AI Agents
A working book on AI agents in 12 chapters and 9 projects: from model, tools and loop to narrow permissions, ledgers that survive, reviewers that push back, prompt-injection defence, cost metering, monitoring and release with human approval.
- Who it is for
- Builders, analysts and team leads who have finished AI Foundations and want to design, run or oversee agents that use tools. Technical chapters use short code examples; every chapter can still be read without writing code.
36 lessons, about 190 minutes
Start this levelPhase 1: Agent basics
Module 1
Chapter 1: What an agent is
An agent as a model, tools and a loop pursuing one goal; how it differs from a chatbot and a fixed workflow; and how to choose its level of autonomy.
Module 2
Chapter 2: Setting up the workspace
A sandbox where an agent may be wrong, environment variables set explicitly, a checkable instruction file and first permission rules.
Module 3
Chapter 3: The first agent
A first agent that only reads, with turn, budget and output limits from the start, then how to read its trace to tell claims from evidence.
Module 4
Project: a folder summariser agent
Build a read-only agent that summarises one test folder with file citations, then check every citation.
Module 5
Project: a team instruction file
Build an instruction file, permission rules and a list of steps an agent never runs alone, then test them with five tasks.
Module 6
Project: an agent task test set
Build a ten-task test set with expected outcomes, then run it before every change to the prompt, tools or model.
Phase 2: Building agents that work
Module 7
Chapter 4: Tools and MCP
Designing tools an agent uses correctly, results and errors a model can understand, write tools that are safe to repeat, and how to assess third-party MCP servers.
Module 8
Chapter 5: Memory and ledgers
Tell one session's context from memory outside it, keep memory notes dated and marked when stale, and use a progress ledger so agent programmes survive broken sessions.
Module 9
Chapter 6: Skills
Package repeated work procedures as skills loaded when needed, write descriptions that trigger at the right moment, and test skills with a task set before sharing.
Module 10
Chapter 7: Agent loops
Set checkable stop conditions, a fixed stop list when the agent gets autonomy, and fix loops driven by tests that prove the change.
Module 11
Chapter 8: Many agents and reviewers
Splitting a programme across subagents with their own lanes and checkouts, evidence-based hand-backs, and a separate reviewer tasked with pushing back.
Module 12
Project: a read-only MCP server
Build a small MCP server with two read tools for one notes folder, then test it directly, through an agent, and with a note hiding instructions.
Module 13
Project: a programme with a ledger
Run five agent tasks with a ledger, stop the session on purpose midway, then prove a new session can resume from the ledger and version history alone.
Module 14
Project: writer and reviewer
Run one real task with a separate writer agent and reviewer agent, then close every finding with evidence.
Phase 3: Agents in production
Module 15
Chapter 9: Permission gates and secrets
Permissions as narrow as the task, a limited environment, secrets the agent never sees, and a case study of a planned command agent design.
Module 16
Chapter 10: Prompt-injection defence
Indirect injection paths, the three danger conditions, data fences, and layered defences that mostly sit outside the model.
Module 17
Chapter 11: Cost and monitoring
Count one agent task's cost, record every AI call against its trigger through one meter door, and monitor steps and processes that can die silently.
Module 18
Chapter 12: Releasing an agent
Ship agent features with the flag off, widen in steps, prepare the way back, and let merge, push and release wait for the owner behind a written gate.
Module 19
Project: an agent with gates
Build an agent for one work tool with three registered actions, previews, typed confirmation and a step log.
Module 20
Project: injection attack test
Attack your own agent with eight planted injections, judge from execution traces, patch with defence layers, then test again.
Module 21
Project: a metered release
Prepare one agent feature for other people: a feature flag, a dry-run meter, a heartbeat, a rollback plan and a release gate the owner approves.
Final assessment
10 scenarios from every module; pass with 8 right. It opens once you have finished every module.