Training / Azure AI Foundations — Agents, Retrieval & Production
Azure AI Foundations — Agents, Retrieval & Production
Build a grounded AI agent on Azure and take it to production: retrieval you tune, tools you define once, evaluation, identity and document-level security.
By the end of eighteen hours, every participant has built a production-ready AI agent on Microsoft Foundry: grounded in a document corpus through an Azure AI Search index using agentic retrieval, extended with a reusable toolbox, evaluated and traced end to end, secured with its own Entra identity and document-level access filters, and published to a stable versioned endpoint. It does several things a low-code platform structurally cannot: retrieval the participant tuned themselves, a model they chose deliberately, access filters that give two users different answers from the same index, and a defence against a document carrying a prompt injection.
Developers new to AI application work who need the whole production path, not a demo that works once.
Teams who have outgrown low-code agents and need to know which ceiling they hit and what replaces it.
Solution architects who must settle retrieval, identity and cost before a project commits to a shape.
Data and platform engineers responsible for index design, chunking and retrieval tuning.
Low-code agents for Microsoft 365. Our Copilot Studio courses cover those, starting with Foundations — Standard Harness.
Fine-tuning or training custom models. This course grounds a model in your data, which solves a different problem.
Voice agents and real-time speech, which are named here but not taught.
Multi-agent orchestration at scale — that comes in a later Azure AI course, in preparation.
Azure administration, networking or identity as disciplines, or Azure AI Search end to end. They are used here at the depth an agent needs.
- Duration
- 3 days or 9 sessions × 2 hours
- Level
- Foundation · Code-first
- Delivery
- Remote or on-site
- Group size
- Up to 10
- Price
- Priced per team — ask us
We reply within one working day. No sales sequence.
What your team will be able to do
Outcomes, not topics covered.
By the end, each of these is something they have done rather than watched.
Explain when a code-first platform is the right choice, and when a low-code agent platform remains the better answer.
Describe the four ways to build an agent on the platform and choose correctly between them.
Create and deploy a prompt agent through both the portal and the SDK.
Design a chunking strategy, build an Azure AI Search index, and ground an agent in it with citations you can verify.
Implement agentic retrieval with query planning, parallel subqueries and semantic ranking.
Diagnose a hallucination back to its cause in the retrieval pipeline rather than blaming the model.
Build a toolbox and reuse it across agents, with the right authentication mode for each tool.
Decide when a prompt agent must become a hosted agent, and what changes when it does.
Build evaluation sets, run them, and read an end-to-end trace.
Give an agent its own identity and apply document-level security filters to retrieval.
Defend an agent against prompt injection carried in retrieved content.
Model the cost and latency of a retrieval-augmented application.
What's covered
16 modules, a little under half of it building.
Labs run in your own Azure subscription, and one agent is taken from a first prompt to a secured, published endpoint across all three days. More time goes on chunking, indexing and retrieval than on model choice, because that is where answer quality is decided. Model deployment quota takes days to raise, so it is checked when the course is booked. We publish what each day produces rather than a minute-by-minute timetable.
Provisional — modules may change before the first run.
Day 1
Ground
Leaves with: An agent that answers from your own documents, with sources you can check
-
01
Where code-first begins
The four ceilings that push a project past low-code agents, and the honest case for staying on one.
-
02
The Foundry platform
Projects, models and the agent runtime, the four ways to build an agent, and what the platform manages for you.
-
03
Lab 1: Your first prompt agent
One agent defined in the portal and again in code, two models compared on the same task, and a call from application code.
Hands-on -
04
Retrieval fundamentals
Retrieve, augment, generate; embeddings and hybrid search; and chunking, the decision everything else inherits.
-
05
Lab 2: Build the index and ground the agent
A chunked corpus, an Azure AI Search index, a grounded agent with verifiable citations, and one re-chunk measured.
Hands-on
Day 2
Build
Leaves with: Retrieval that handles follow-ups and multi-document questions, and tools defined once
-
06
Agentic retrieval
Query planning from conversation history, parallel subqueries, semantic ranking, and when it earns its cost.
-
07
Lab 3: Agentic retrieval and citation quality
Multi-turn and multi-document questions answered, and a planted hallucination traced to a corpus gap rather than the model.
Hands-on -
08
Tools and toolboxes
Built-in tools, toolboxes defined once and reused, and choosing between agent identity and on-behalf-of authentication.
-
09
Lab 4: Build a toolbox
A toolbox with a custom tool and on-behalf-of authentication, shared by two agents and versioned.
Hands-on -
10
Hosted agents and code-first
When a prompt agent must become your own code, what the platform still provides, and the lighter Responses API.
Day 3
Prove and ship
Leaves with: An evaluated, secured agent on a versioned endpoint, with a cost per request
-
11
Evaluation and tracing
Evaluation sets and synthetic cases, reading a trace from output back to retrieval, and the optimiser as a suggestion engine.
-
12
Lab 5: Evaluate and trace
An evaluation set with a recorded baseline, and one failure diagnosed from a trace and fixed by a single change.
Hands-on -
13
Identity, security and grounding safety
An identity per agent, document-level security filters, and prompt injection carried in retrieved documents.
-
14
Lab 6: Secure and publish
Two users get different answers from one index, a planted injection is defeated, and the agent ships to a versioned endpoint.
Hands-on -
15
Cost, latency and model economics
Where the tokens go, why retrieval is not free, and a monthly cost model with retrieval overhead shown separately.
-
16
Capstone and the pathway
The agent presented retrieval design first, and the first morning's code-first decision revisited.
Upcoming batches
No public batch scheduled yet.
Teams book this course on their own dates. Joining as an individual? Tell us and we'll let you know when the next batch opens.
Before the first day
What we need from you.
Prerequisites
- Comfort reading Python or C#. The labs supply scaffolds to complete; no production code and no AI or machine learning background.
- Familiarity with Azure (resource groups, identity, the portal) shortens Day 1 but is not required.
- An Azure subscription with permission to create a Microsoft Foundry project and an Azure AI Search resource.
- Model deployment quota in your region. Increases take days, so raise it when the course is booked.
- Permission to create an app registration, for Lab 4, and an Application Insights resource.
- The SDK installed locally, confirmed by a supplied hello-world script.
- Every participant completes an access check 48 hours before day one.
We send a checklist a fortnight ahead and will talk to your platform team directly if access needs arranging.
Next step
Tell us about your team.
Roughly how many people, what they build today, and when you'd want to run it. We come back with dates and a quote.
Check availability