Agentic AI Foundations for Malaysian Graduates

Agent concepts and use cases

• Distinguish a language model, chat interface, workflow and tool-using agent; capability boundaries overlap and are not fixed by marketing labels.
• Illustrative research, information-organisation and personal-assistant tasks; distinguish drafting from authorised action.
• Discuss work-related uses without stale labour-market statistics or promises that AI solves job scarcity.

Algorithmic and systems thinking

• Break a task into steps, rules and decision points using flowcharts or simple pseudocode.
• Identify data, tools, people, outputs, dependencies and feedback loops.
• Separate appropriate automation from human judgement; define missing-information handling and a stopping condition.

Prompt foundations

• State the goal, context, output format, criteria and constraints.
• Compare vague and specific requests and practise iterative revision.
• Inspect an agent's visible plan/actions as execution evidence, not a faithful transcript of private reasoning.

Guided activities and tool awareness

• Map an event-planning or survey-information task; practise a simple prompt with an available chatbot.
• Observe one prepared agent research demonstration in a learning environment.
• Introduce hosted agent tools such as Manus, AutoGPT deployment options and simple diagramming tools without assuming free access.

Week-one mini-project

• Design an illustrative virtual assistant's narrow task specification.
• Define inputs, permitted tools, expected outputs, checkpoints and explicit exclusions.
• Present the process map and identify assumptions needing verification.

Clear goals and iterative review

• Use context, roles, limits, criteria and examples; treat SMART-style templates as an organising aid, not a proven universal recipe.
• Plan, run, check and refine a limited task; keep prompts focused.
• Compare output quality, sources, completeness and failure cases rather than accept fluent text.

Current tools and access models

• Compare an available hosted agent with a prepared self-hosted or framework-based demonstration.
• Distinguish AutoGPT Classic from the current Platform; licence, models, setup and costs vary by component.
• Manus capabilities are evaluated in a selected learning task, not assumed reliable for every web action.
• CrewAI previews role-based coordination; historical Cognosys/AgentGPT examples are not mandatory unavailable-platform dependencies.
• Review plans, credits, model costs, permissions and data handling before using a service.

Execution workflow

• Define a goal, choose an appropriate tool, configure boundaries and observe progress.
• Use approved sources, respect login/paywall/access restrictions and stop when necessary.
• Review results and intervene on loops, errors, unsupported claims or unexpected actions.
• Do not share provider keys among students; use authorised learning accounts or a controlled demonstration.

Prompt and comparison labs

• Write a bounded information-gathering prompt and test or simulate it.
• Observe selected browser/data-to-table tasks with synthetic records and no real-account changes.
• Compare tools using observed evidence rather than unsupported performance rankings.

Week-two mini-project

• Research a fictional small business's competitors using approved public or prepared information.
• Separate observations, cited evidence, assumptions and sentiment interpretations.
• Present the agent/task setup, findings, limitations and revisions without claiming measured time savings unless actually recorded.

Process mapping and control

• Map an illustrative customer-support, inventory, event or onboarding process with human/agent swimlanes.
• Define data hand-offs, review points, bounded retries, cost/time limits and failure escalation.
• Keep consequential hiring, ordering, financial and communication decisions with authorised people.

Single versus multiple agents

• Compare one tool-using agent with specialised roles such as researcher, summariser and reviewer.
• Inspect a prepared CrewAI example and its distinction between Crews and controlled Flows.
• Use AutoGen/other frameworks as conceptual comparisons, subject to their current support/version requirements.
• Multiple agents can add coordination cost and propagate errors; they are not automatically better or independent verification.

Tools, knowledge and no-code integration

• Explain approved API/tool access, source retrieval and permission boundaries.
• Compare supported connectors/actions and no-code platforms such as Zapier or Make without relying on obsolete ChatGPT Plugins exercises.
• Distinguish retrieving approved information from training a model on private data.
• Review human approval for draft replies or updates; use a manual gate where a native mechanism is not supported.

Activities and a light technical view

• Role-play data gathering, reading and summarising to expose hand-off failures.
• Draw a workflow with at least one explicit human checkpoint and test a failure scenario.
• Observe a prepared CrewAI or equivalent example; pseudocode illustrates bounded control, not a coding prerequisite.
• Keep a simple version record of prompts, assumptions, test inputs and results.

Week-three design project

• Describe the current process, proposed roles, data flow and permitted actions.
• Use synthetic cases and approved sources rather than scrape private profiles or collect sensitive hiring information.
• Present a blueprint with validation, access, governance and feasibility questions; do not claim a full implementation.

Illustrative business applications

• Compare marketing drafts, customer-support summaries, research briefs, operational assistance and creative ideation.
• Analyse data quality, evidence, professional review and organisational permissions for each scenario.
• Estimate costs and benefits using explicit assumptions; do not treat an agent as a legally accountable co-founder or guaranteed substitute for staff.

Deployment and continuing review

• Demonstrate or simulate deployment of one selected prototype to a controlled learning environment.
• Explain scheduling, triggers and reporting conceptually; scheduled/external actions need separate explicit permission in real use.
• Review monitoring, feedback, change control and when to pause or return a task to a person.
• Assess current tools by their actual documentation and licence rather than dated 2024–2025 hype or future predictions.

Portfolio and responsible communication

• Document what was actually built, designed or observed and the learner's specific contribution.
• Separate simulated examples from operational deployments; do not invent client work, percentages, achievements or credentials.
• Discuss ethical/social trade-offs and sources of further learning without guaranteed career advancement.

Final action-plan options

• Career/application track: identify two or three suitable tasks, tools, constraints, proposed pilot steps and further support needed.
• Entrepreneurial track: describe an illustrative service idea, tool costs, feasibility and customer validation assumptions.
• Choose one bounded prototype or process proposal and explain evidence, limitations, approval points and next tests.
• Present the demonstration/action plan for peer review and refine it using feedback.