WEEK 01
v1.0
Foundation
3 sessions
Analysis Copilot
Build the agent loop by hand so that no framework is ever magic. By Friday you have an agent that plans its own steps and defends its answers.
- BUILDThe think → act → observe loop, written from scratch
- BUILDTool calling with schemas, validation, and what to do when a tool lies
- SHIPAn agent that answers data questions with citations and a cost trail
- DEEP DIVEWhy most agents fail on step three, and how to see it happening
agent looptool callingplanningtracing
WEEK 02
v2.0
Retrieval
3 sessions
Knowledge Navigator
Retrieval that actually retrieves. Most RAG demos work because the corpus is tiny — this week is about what happens when it isn't.
- BUILDChunking as a design decision, not a default parameter
- BUILDHybrid search and reranking that survives real documents
- SHIPA navigator that knows when to say “I don't know”
- DEEP DIVEThe retrieval failure nobody logs, and how to measure it
RAGvector searchrerankingabstention
WEEK 03
v3.0
Orchestration
3 sessions
AI Engagement Team
Several agents that divide work, disagree, and still reach a decision — without looping forever or quietly burning your budget.
- BUILDHandoffs and arbitration between specialised agents
- BUILDTermination conditions, retries and cost ceilings
- SHIPA multi-agent workflow that completes reliably, not occasionally
- DEEP DIVEWhen multi-agent is genuinely the wrong answer
multi-agentorchestrationhandoffscost control
WEEK 04
v4.0
Production
3 sessions
Enterprise Action Assistant
The week every other course skips: the one where the agent touches real systems, and someone senior asks whether it's safe.
- BUILDPermissions, audit trails and a kill switch
- BUILDAn evaluation harness that runs before every change
- SHIPAn assistant that passes a production design review
- DEEP DIVEObservability, governance, and defending the design to your CISO
guardrailsevalsobservabilitygovernanceMCP
