CoderPush × Sun Group

AI-Native Engineering
for the Modern Enterprise

How AI agents are transforming the way software gets built

Harley Trung · CEO & Co-founder, CoderPush

Today's Agenda

The evolution of AI-assisted development
What are coding agents & how they work
From specs to working software
Two pillars for your team
- Great specs — for BAs & PMs
- Skills & agents — for developers
What this means for Sun Group

Part 1

What is Agentic Coding?

The Evolution of AI & Coding

Code Completion

1

Code Completion

2

Code Generation

3

Agentic Coding

Autocomplete

The Evolution of AI & Coding

Code Generation

✓

Code Completion

2

Code Generation

3

Agentic Coding

LLM-Powered Code Generation

The Evolution of AI & Coding

Agentic Coding

✓

Code Completion

✓

Code Generation

3

Agentic Coding

Autonomous AI in Your Workspace

Coding Agents

IDE

Copilot in VS Code
Cursor
Windsurf
Jetbrains

CLI

Copilot CLI
Claude Code
Codex
Kiro (AWS)

Web

GitHub Copilot
ChatGPT's Codex
Claude
Devin.ai

Custom

Copilot SDK
OpenAI Agents SDK
LangChain's Deep Agents SDK
Claude Agents SDK

Core Concept

What is an AI Agent?

A widely accepted definition today:

An AI Agent is an LLM that calls tools in a loop to achieve a goal.

The Agentic Loop

How coding agents think, act, and iterate

Source: code.claude.com/docs/en/how-claude-code-works

Part 2

Use Case: Specs to Software

Rebuilding a Sun Group page with AI agents

Overview

The Full Pipeline

📄

Documents
docx / xlsx

→

📋

Markdown
Spec

→

✅

Review
human

→

🎫

GitHub
Issues

→

🤖

Agent
Codes

→

🔍

PR
Review
human

→

🚀

Deploy

This pipeline is similar whether you're building a project showcase, an internal tool, or a customer-facing application.

Live Demo — Act 1

From Documents to Specs & Issues

Your BAs write Word docs and Excel sheets. Watch what happens when an AI agent processes them.

What you will see

Input: one or more .docx / .xlsx files
AI extracts, normalizes, and decomposes into focused specs
Human reviews and approves
AI generates GitHub / Jira Issues from approved specs

📄 → 📋 → 🎫

Minutes, not days.
Consistent format every time.

Live Demo — Act 1

Decompose, Don't Just Convert

The relationship from input documents to specs is never 1:1. Specs should be organized by delivery boundary, not by source document.

Input

📄

SunGroup-BRD.docx

→

Output — 4 focused specs

spec-project-discovery-behavior.md
spec-project-data-contract.md
spec-filtering-and-sorting-logic.md
spec-quality-and-acceptance.md

Every requirement traces back to its source via a traceability index. Not spec-v2-final. Not requirements-copy-3. Names that stay stable and mean something to every role on the team.

Live Demo — Act 2

From Issues to Working Code

An agent picks up an issue, reads the spec, and builds the feature — with skills to accelerate it.

Without Skills

The agent writes generic code. It might use outdated patterns, miss best practices, or produce something that needs heavy refactoring.

🐌

More iteration cycles

With Curated Skills

The agent reads skill files first — learning your team's patterns, tech stack, and standards. It produces code that matches your architecture from the start.

⚡

Production-ready faster

Live Demo — Result

Sun Group Project Showcase

[ LIVE DEMO ]

Switch to browser to show the finished application

Best Practices

✦ · ✦ · ✦ · ✦ · ✦

Best Practice #1

Context

1

skill.md

2

agents.md

3

Prompting

Curated Skills

Curated Skills (beige) improve performance by +16.2pp on average; self-generated Skills (amber) provide negligible or negative benefit.

Figure 1. Agent architecture stack and resolution rates across 7 agent-model configurations on 84 tasks. — arxiv.org/html/2602.12670v1

Best Practice #1

Context

✓

skill.md

2

agents.md

3

Prompting

agents.md

Developer-provided files only marginally improve performance (+4% on average), while LLM-generated context files have a small negative effect (−3%). Context files increase costs by over 20%.

These observations are robust across different LLMs and prompts. — arxiv.org/html/2602.11988v1

Best Practice #1

Context

✓

skill.md

✓

agents.md

3

Prompting

Prompting: Explicit is better than implicit

Reference specific files to show example code
Describe the expected output format
Break complex tasks into smaller, focused prompts
Include constraints and edge cases up front

The more precise your prompt, the better the agent's first attempt — saving iteration cycles.

Best Practice #2

Verification

Unit Tests

Validate individual functions and modules produced by the agent.

Traces

Inspect agent reasoning and tool calls to understand its decisions.

Functionality Tests

End-to-end checks that the feature works as intended.

Best Practice #2

Real Browser Validation

Run the shipped app in a real browser, exercise key flows, and document the result in issue #1. That gives stakeholders a concrete artifact even when there is no separate PR.

Final deck step: attach this screenshot to the GitHub issue, then capture the issue page itself if you want the slide to show the GitHub artifact instead of the raw browser view.

Best Practice #3

Version Control

Commit often — create checkpoints as the agent works
Maintain documentation — keep READMEs and docs in sync
Be mindful about more PRs created — review agent-generated code carefully

Part 3

What This Means
for Sun Group

The Impact

3–5×

Faster prototyping
from spec to working code

80%

Reduction in spec-to-issue
handoff time

∞

Scales with your team —
agents don't burn out

These aren't hypothetical. This is what we see across our engineering teams today.

Two Pillars for Your Team

Pillar 1: AI-Ready Specs

For BAs, PMs, and product teams

Structured requirement templates
Document-to-spec conversion
Automated issue generation
Quality gates before handoff

Pillar 2: AI-Native Development

For developers and tech leads

Curated skill libraries
Agent configuration and setup
AI-assisted code review
20% time for AI training

What CoderPush Delivers

🎓

Training Playbook

Hands-on workshops for BAs, PMs, and developers. Your team learns by building real features with AI agents.

🛠️

Custom Skill Library

Curated skills tailored to your codebase, architecture, and workflows. Agents that know your standards.

📐

Spec Framework

Templates that turn your existing document workflows into agent-ready specifications.

Let's Build Together

The best time to adopt AI-native engineering was yesterday.
The second best time is today.

What We Covered

Code completion → generation → agentic coding
Live demo: documents to deployed application
Best practices: context, verification, version control
Two pillars: AI-ready specs + AI-native development

Harley Trung

CEO & Co-founder, CoderPush · coderpush.com

✦ · ✦ · ✦ · ✦ · ✦

AI-Native Engineeringfor the Modern Enterprise

Today's Agenda

What is Agentic Coding?

Code Completion

Autocomplete

Code Generation

LLM-Powered Code Generation

Agentic Coding

Autonomous AI in Your Workspace

Coding Agents

IDE

CLI

Web

Custom

What is an AI Agent?

The Agentic Loop

Use Case: Specs to Software

The Full Pipeline

From Documents to Specs & Issues

What you will see

Decompose, Don't Just Convert

From Issues to Working Code

Without Skills

With Curated Skills

Sun Group Project Showcase

Best Practices

Context

Curated Skills

Context

agents.md

Context

Prompting: Explicit is better than implicit

Verification

Unit Tests

Traces

Functionality Tests

Real Browser Validation

Version Control

What This Meansfor Sun Group

The Impact

Two Pillars for Your Team

Pillar 1: AI-Ready Specs

Pillar 2: AI-Native Development

What CoderPush Delivers

Training Playbook

Custom Skill Library

Spec Framework

Let's Build Together

What We Covered

AI-Native Engineering
for the Modern Enterprise

What This Means
for Sun Group