Agent Foundations
What an AI agent really is, how it uses tools and context, and how to make one dependable enough for high-stakes work.
2 modules · 6 lessons~18 hoursUpdated 2026-10-30
- Lessons6
- Quizzes6
- Sources36
- Of study~18h
About this course
About this course
The second course in the LLM and Agent Foundations path. It builds the core mental model from first principles: an LLM call, then an LLM with tools, then a loop that decides its own next step. It then covers the engineering that makes agents dependable: tool design, context engineering for long documents, structured and grounded outputs, the standard workflow patterns, failure modes and a first small eval. Every idea is worked through on a realistic document-heavy workflow and practised in Claude Code and claude.ai on your Claude subscription.
About the running example
The examples follow a fictional property-law (conveyancing) firm in England and Wales: its agents read title registers, leases and property searches. It's a case study, chosen because its documents are long, structured and high-stakes. You don't need a legal background, and the same patterns apply to contracts, insurance claims, invoices, support tickets or patient records. In the lessons, a matter is one client's case and a solicitor is the lawyer who signs work off.
What you'll learn.
- 1Explain the difference between an LLM call, an augmented LLM, a workflow and an agent, and decide which one a real task needs
- 2Design tools for an agent, including schemas, error messages and mistake-proofing, and connect one to Claude Code as a script or MCP server
- 3Engineer the context an agent sees when it works through long documents such as contracts, reports and forms
- 4Produce schema-validated JSON with evidence quotes, and check those quotes against the source in code
- 5Map chaining, routing, parallelisation, orchestrator-workers and evaluator-optimizer onto the steps of a real business workflow
- 6Identify agent failure modes, including prompt injection from documents, and measure an agent with a 10-case eval
2 modules, 6 lessons.
7h 30m of lessons
Module 1 · 3 lessons · 3h 40m
What agents are
The mental model: from a single model call to a loop with tools, and the two things that decide how well it works: the tools it can use and the context it sees.
Module 2 · 3 lessons · 3h 50m
Making agents dependable
Typed, evidence-backed outputs, the standard patterns for composing model calls, and the failure modes you must design and test for.
A document review agent in Claude Code
Build a Claude Code agent that reviews ten synthetic official records with a subagent, a skill and your own lookup tool, returns schema-validated JSON with checked evidence quotes, and is measured by a 10-case eval. Runs entirely on your Claude subscription.