Using Agentic AI Coding

AI-Native Technologist

This workshop is part of our AI-Native Technologist learning path. You see the full path here

AI-Native Technologist Learning Path

Learning Objectives

• Describe the concept of Agentic AI and its applications
• Differentiate between Agentic AI and traditional AI tools
• How to use Agentic AI, without diminishing human ownership and responsibility

Pre-work

Workshop

Reflect

Ask trainees to reflect on their experience with AI so far

• How have they used it well?
• What have they found frustrating?
• What have they found surprising?
• What have they found disappointing?

Discuss - Understanding Agentic AI

Ask trainees to share their definitions of Agentic AI.

Collect core traits on a shared board:

• Autonomous decision-making
• Goal / state orientation
• Multi-step planning
• Context persistence
• Self-monitoring & adjustment

Differentiate an agent from a passive assistant: initiative, planning, action sequencing, boundary negotiation.

Live Example - Assistant vs Agent

Demonstrate two approaches to: "Organize project files by feature area and add missing README stubs."

Show assistant pattern: repeated Q&A, manual steering.

Show agent pattern: scans structure, proposes taxonomy, outlines plan, requests approval, drafts README templates.

Highlight that agents propose, plan, and (within limits) act; assistants mostly react.

Prompt Design - Small prompts

Explain that small, specific prompts guide AI behavior for limited tasks and reduce ambiguity.

Explain that working with AI is like working with an eager, talented 15-year-old intern: all of the ability but none of the judgment; clear, focused instructions yield best results

Live code the example below to show how short, specific instructions get the best results from an AI model.

• Add Feature 1: Add prices to items
• Add Feature 2: Calculate total cost
• Add Feature 3: Generate shopping basket summary

def shopping_basket(items):
    """
    Create a shopping basket summary.

    Parameters:
    items (list of str): List of item names to include in the basket.

    Returns:
    dict: A dictionary with item names as keys and quantities as values.
    """
    basket = {}
    for item in items:
        if item in basket:
            basket[item] += 1
        else:
            basket[item] = 1
    return basket

shopping_basket(['apple', 'banana', 'apple', 'orange', 'banana', 'apple'])

Prompt Design

Highlight that, as always, context is key when working with LLMs.

Show how in the agent view you can pull in all of the context you need to be able to do the role.

Show how you will get poor results when you don't give enough context

Present a weak instruction: "Improve code quality."

Refine it live into a structured agent directive:

• GOAL: Raise maintainability of Python services (no functional changes).
• BOUNDARIES: No external API calls; no deletion of files; propose changes via patch diff.
• CHECK-INS: After inventory + after first refactor suggestion.
• SUCCESS: Reduced cyclomatic complexity (≤10 per function) & added missing docstrings.
• ESCALATE: Any uncertainty about intent.

Explain that explicit boundaries + checkpoints prevent silent scope drift.

Remind trainees of the Capabilities versus Token Burn discussion from week one. An agent loop re-sends its context with every turn, so token burn compounds: a conversation that touches many files can consume far more tokens than the same task done in a single chat. This is another reason to keep prompts small and boundaries tight, and to consider whether the whole job needs a frontier model or whether a cheaper tier will do.

Guided Exercise - Designing a Code Quality Agent

Form small groups (3-4).

Provide scenario: Large codebase suffering from style inconsistency and duplicated utility functions.

Instruct groups to draft an agent spec (10-12 min): scope, forbidden actions, required data, human review triggers.

Agent Design Acceptance Criteria

Acceptance criteria

• Single clear mission statement
• Explicit forbidden actions list
• At least 2 human approval checkpoints
• Logging/audit plan (diff summary + rationale)
• Escalation rule for uncertainty / ambiguity

Collect specs and briefly compare boundary clarity.

Copilot Instructions

Highlight that the .copilot-instructions.md is key for working with Copilot Agents, as it allows you to set clear boundaries and instructions for the agent to follow.

Task the trainees with using AI to generate their own .copilot-instructions.md file for this week's coursework

Ask the trainees about the instructions they generated, and how they think it will affect the behavior of the agent.

Ask the trainees about how important it would be to keep this file up to date, and what the risks would be of it being out of date.

Human Ownership & Responsibility

Emphasize three principles:

• Human final decision
• Clear constraints
• Accountability never delegated

Discuss examples: security scan false positives, risky refactors, dependency upgrades, architectural rewrites.

Warn that silent agent changes without review create integrity & trust risks.

Explain that we'll be looking at that next week

Conclusion

Reinforce that Agentic AI pays off only when ownership, constraints, and review loops are explicit.

Remind trainees they remain answerable for outcomes produced by agents.

Encourage trainees to plan their prompts carefully before enabling autonomy.