Audience
- Product and operations professionals designing AI-assisted services.
- Developers or automation practitioners adding tool-using assistants to applications.
- Technical consultants who need to specify and evaluate an agent workflow.
AI & technology
AI Agents Design
Design useful agents with clear responsibilities, tools and human oversight.
Learning with a practical purpose.
Design AI agents around a useful responsibility rather than an open-ended instruction to do everything. This module develops a small assistant that can interpret a request, consult approved information and use a limited set of tools. You will define what the agent may do, what it must ask about and when it should hand work to a person.
The practical focus is the relationship between instructions, tools, state and evaluation. You will compare an agent with a fixed workflow, inspect the steps it takes and test how it behaves when information is missing or contradictory. By the end, you will have a scoped prototype and an evidence-based explanation of where agent behaviour helps the chosen business task.
Live onlineTeaching through Kent Business College, UK
2 hoursOne live session each teaching week
14 sessionsA semester for each module
28 live hoursPer module, plus independent study
The outline below is a 14-session online module, with one two-hour live session each teaching week. Reading weeks, breaks and additional activities follow the agreed cohort calendar. The cohort syllabus, full route, dates, fees, independent study and assessment requirements are confirmed before enrolment.
Who this module is designed for.
Starting knowledge and preparation
- Basic understanding of APIs, structured data and business workflows.
- Some Python or JavaScript experience, or equivalent practical automation experience.
- Familiarity with using a language-model interface is helpful.
What you will work towards.
- Define an agent’s purpose, boundaries and success measures for a specific workflow.
- Design clear tool contracts and separate read-only actions from consequential changes.
- Manage task state and selected memory without retaining unnecessary information.
- Evaluate tool selection, task completion and handover using realistic scenarios.
- Build a limited agent prototype with observable steps and human checkpoints.
Explore the 14-session learning plan.
Explore the focus and practical output of each teaching session. Open individual sessions below, or expand the complete plan.
Session 01 When an agent is useful
Compare a fixed workflow, a simple assistant and an agent against the same business requirement.
Practical output: An approach-selection decision brief.
Session 02 Responsibilities and task boundaries
Identify permitted tasks, excluded actions and the evidence needed before an agent can claim completion.
Practical output: An agent responsibility specification.
Session 03 Instructions and operating context
Write instructions that separate role, procedure, constraints and task-specific information without unnecessary repetition.
Practical output: A structured instruction set.
Session 04 Designing tools the agent can use
Specify names, inputs, outputs and error messages for a small collection of business tools.
Practical output: A tool catalogue and input schemas.
Session 05 Connecting read-only capabilities
Give the agent controlled lookup or calculation tools and inspect whether it selects them appropriately.
Practical output: A read-only tool-using assistant.
Session 06 Reasoning steps and workflow state
Represent task progress explicitly so incomplete work can be recognised, resumed or handed over.
Practical output: A task-state model and execution trace.
Session 07 Short-term context and selected memory
Distinguish current conversation context from information worth retaining across sessions and define retention boundaries.
Practical output: A memory design with update rules.
Session 08 Knowledge access and evidence
Connect an approved knowledge source and require the agent to distinguish retrieved evidence from unsupported assumptions.
Practical output: A source-aware answer workflow.
Session 09 Approvals and consequential tools
Introduce a controlled write action with explicit approval, validation and a record of the outcome.
Practical output: An approval-based action workflow.
Session 10 Failure, ambiguity and handover
Test missing information, unavailable tools and conflicting requests, then define useful recovery and escalation responses.
Practical output: A handover and recovery scenario set.
Session 11 Testing the agent systematically
Create cases that measure tool choice, task completion and adherence to the declared operating boundaries.
Practical output: An agent evaluation suite.
Session 12 Adversarial inputs and instruction boundaries
Explore how untrusted retrieved content can conflict with instructions and test whether protected actions remain restricted.
Practical output: A boundary-testing report.
Session 13 Prototype integration and observation
Combine instructions, tools, state and evaluation in a small application with useful execution records.
Practical output: An integrated agent prototype.
Session 14 Demonstration and operating recommendation
Present normal and difficult scenarios, explain the results and propose a limited supported use of the agent.
Practical output: A final demonstration and operating proposal.
Apply the learning
A scoped service-support agent
Build an assistant that handles a defined category of enquiries, consults approved information and prepares a proposed next action. Include at least one read-only tool, an approval-dependent action and a human handover. Your demonstration must show how the system recognises incomplete information and how its behaviour was evaluated.
Your project evidence can include
- An agent specification, tool contracts and state diagram.
- A runnable prototype with observable execution steps.
- An evaluation report and supported-use recommendation.
Use anonymised, public or suitable synthetic material where workplace information is confidential. The issued assessment brief confirms the required evidence.
Assessment, feedback and certification.
- Design exercises assess task boundaries, tool clarity and memory decisions.
- The prototype is tested against ordinary, ambiguous and failure scenarios.
- A professional discussion examines why the agent approach is justified and where human oversight remains necessary.
Teaching and the relevant course certificate are provided through Kent Business College in the UK. Completion requirements are set out in the module information; attendance alone should not be assumed to meet every assessment requirement.
Questions worth asking.
Will I build a fully autonomous employee?
The project is a narrowly scoped assistant with explicit limits and review points. Broad autonomous responsibility is not an assumed learning outcome.
Does the module require several agents?
No. You first establish a useful single-agent design. Additional agents are considered only when separate responsibilities justify the coordination overhead.
How is this different from RAG?
RAG focuses on retrieving knowledge to inform a response. This module also covers selecting tools, managing task state, approvals and completing a bounded workflow.