Agentic AI Systems Engineering

About this course

Agents promise to automate multi-step work, and most production failures come from giving them more autonomy than the task needs. This course teaches the full design space, from fixed workflows to autonomous agents: tool design and the Model Context Protocol, planning and memory, multi-agent systems, durable execution, and computer-use and coding agents. It insists on measurement and safety: task-level evaluation, trajectory analysis, cost control, sandboxing, least privilege and human oversight, so that learners ship the simplest agentic system that meets the requirement and can account for everything it does.

For: software engineer, ml practitioner, architect.

Learning outcomes

On successful completion you will be able to:

1. 1Explain the design space of agentic systems, from fixed workflows to autonomous agents, including tool use, planning, memory and multi-agent coordination.
2. 2Build tool-using agents with well-designed tool interfaces, including Model Context Protocol servers and clients.
3. 3Design and implement a multi-step agentic system with durable state, error recovery and human approval points.
4. 4Evaluate agent behaviour with task-level benchmarks, trajectory analysis and cost and latency measurements, and choose the simplest architecture that meets the requirement.
5. 5Design guardrails, sandboxing and permission models that bound an agent's autonomy and limit the blast radius of failures and attacks.
6. 6Decide where autonomy is and is not acceptable in a business process, and take responsibility for the oversight, accountability and auditability of agent actions.
7. 7Lead an agentic system from prototype to production, setting quality gates and communicating risk and limitations to stakeholders.

Modules and lessons

1. Week 1Workflows and agents⌄

   • Workflows versus agents: choosing the right autonomy
   • The agent loop: reasoning and actinglab
2. Week 2Tools and the Model Context Protocol⌄

   • Designing tools for agents
   • The Model Context Protocollab
3. Week 3Planning and memory⌄

   • Planning, decomposition and reflection
   • Memory and state for agents
4. Week 4Multi-agent systems and durable execution⌄

   • Multi-agent architecturescase study
   • Durable execution and long-running agentslab
5. Week 5Agents that act on computers and code⌄

   • Computer use and browser agents
   • Coding agents and software-engineering automationlab
6. Week 6Evaluating agents⌄

   • Evaluating agents: tasks, trajectories and outcomes
   • Observability, cost and latency of agentslab
7. Week 7Safety and bounded autonomy⌄

   • Securing agents
   • Human oversight and accountability
8. Week 8Agents in production⌄

   • Agent platforms and build-versus-buy
   • From prototype to production

Assessment

• Quiz (15%)

  Eight weekly quizzes drawn from the lesson quiz banks; best seven count.

• Lab (30%)

  Five graded labs, from a hand-written agent loop to a resumable, instrumented agent with an MCP server.

• Project (35%)

  Build, evaluate and harden an agent for a realistic multi-step business task; deliver code, an evaluation report with trajectories and costs, and a safety case.

• Viva (20%)

  Twenty-five-minute defence of the project's autonomy, permission and oversight decisions before two mentors.

Pass mark 60%. Grades: Distinction, Merit, Pass.