Enterprise AI Agent Engineering

The Problem

Most teams encounter AI agents through demonstrations. Once the demo works, important questions arise: How do we know a change made the system better? What prevents an agent from doing something it should not? How does the agent reach real services and organisational knowledge? This course addresses the tools, checks, evaluations, and controls required to build AI agents that can be measured, trusted, and explained.

Course Approach

Participants work on one running IT-operations agent called Compass and apply each concept to their own organisational services through guided exercises and worksheets.

Learning Outcomes

• Call an AI model directly and read cost and latency from application logs before adopting an AI framework.
• Turn free text into validated, structured and machine-readable data.
• Manage prompts as versioned files.
• Prove that a prompt, model or tool change is better using an evaluation suite whose result can gate a release.
• Measure evaluation pass rate, cost, latency and run-to-run variation.
• Turn an existing service into agent-ready tools and expose them through a simple MCP server.
• Add retrieval over company knowledge with citations and access control.
• Choose between a single prompt, a workflow and an agent based on the problem.
• Evaluate whether to build or adopt an AI framework.
• Keep humans in control of risky actions using budgets and approvals.
• Threat-model an agent system and apply controls such as access-controlled retrieval, approvals and budgets.
• Explain the role of authentication, gateways and audit logs in production environments.
• Produce a ranked recommendation for organisational services: expose read-only first, expose later, or do not expose.

Session 1: The LLM as a System Component

Session 2: Prompt Engineering and Structured Output

Session 3: Evaluation Engineering

Session 4: Agent-Ready Service Design and MCP

Session 5: Knowledge Architecture

Session 6: Workflow Patterns and Framework Landscape

Session 7: Autonomous Agents and Guardrails

Session 8: MCP in Production, Governance and Wrap-Up

Delivery Format

• Virtual and instructor-led training.
• Eight sessions of 5 hours (300 minutes) each.
• Each session includes two breaks.
• Participants work individually; there is no group work or peer review.

Live Coding

• The instructor builds the Compass project one concept at a time.
• Participants type and run small sections of code alongside the instructor.
• Participants can use the next-step scripts if they fall behind.
• Practice is supported through a Git repository containing simple runnable scripts for each session.

Session Structure

• Warm-up poll.
• Concept segments with code.
• Guided coding lab.
• Short threat demonstration.
• Architects clinic.
• Session wrap-up.

Homework and Assessment

• Light homework after each session.
• Re-run the session code.
• Try one easy extension.
• Complete a 10-question self-check quiz.
• Complete a worksheet section.
• Low-stakes assessment with no grades.

Compass

Participants build a working, tested learning project called Compass. The project includes a simple MCP server and an agent that uses it.

Production Scope

The course does not build a production system. Production topics including OAuth, gateways and audit logs are taught as concepts so participants understand what should be added and when.

Primary Model

The main laboratory model is Claude.

Other Providers and Frameworks

Other AI providers are compared on paper or shown through demonstrations. Frameworks such as LangGraph and Microsoft Agent Framework are covered at an awareness level.

Do I need to know AI or machine learning?

No. Participants need working proficiency in one programming language, comfort reading intermediate Python, and familiarity with REST/JSON, Git and the command line. The course starts from a plain model call.

Is the course suitable for architects who do not code much?

Yes. Architects can run the provided code and focus on design decisions, decision records and worksheets while still completing the same working system.

How much time does the course take?

Each of the eight sessions is 5 hours (300 minutes), including two breaks. Participants should also expect light homework after each session.

Is there group work or peer review?

No. Participants work individually on their own copy of the code. Assessment is low-stakes and self-run, with no grades and no peer review.

What do I need to participate?

Participants need a laptop with recent Python 3 and VS Code, Git and a GitHub account, Node.js, Claude Code or access to its demonstration, and an API key with a spend limit for the main model provider. A second provider key is optional.

Will I build a production system?

No. Participants build a working and tested learning project called Compass, including a simple MCP server and an agent that uses it. Production topics such as OAuth, gateways and audit logs are taught as concepts rather than implemented.

Which AI vendor does the course use?

The main lab model is Claude. Other providers are compared on paper or demonstrated, while frameworks such as LangGraph and Microsoft Agent Framework are covered at awareness level.