Practical Agentic AI in Software Engineering

Talk presented at The National University of Singapore Instutite of Systems Science (NUS-ISS) Learning Festival 2025.

Companion code: https://github.com/uzyn/nusiss-learningfest-2025

Transcript

1. Practical Agentic AI in Software Development chua@uzyn.com U-Zyn Chua https://uzyn.com · LinkedIn: uzynchua
2. About me • Independent software engineer • NUS-ISS Adjunct Lecturer • teaching Architecting Agentic AI Solutions • Singapore's first Smart Nation Fellow • Ex-CTO & founder of technology startups. https://uzyn.com
 chua@uzyn.com
 LinkedIn: U-Zyn Chua
3. Image by Thomas Chan. Free to use under Unsplash License. Quick survey
4. Agenda • What is Software Development? • Software development with Agentic AI • Demo! 🙏 • Context engineering • Tips
5. Software Development https://datarob.com/essentials-software-development-life-cycle/
6. Software Development Lifecycle https://www.ibm.com/think/topics/sdlc 1. Planning 2. Analysis 3. Design 4. Coding 5. Testing 6. Deployment 7. Maintenance
7. Software Dev: Planning https://www.ibm.com/think/topics/sdlc • Define goals • Define requirements • Identify scope Output: 
 Project plan document
8. Software Dev: Analysis https://www.ibm.com/think/topics/sdlc • Review the plan • Research • Refine Output: 
 Product requirement document (PRD)
9. Software Dev: Design https://www.ibm.com/think/topics/sdlc • Define project architecture • Software stack • Database design • UI Output: 
 Software design document (SDD)
10. Software Dev: Coding https://www.ibm.com/think/topics/sdlc Implement, basically. Output: 
 Functional software prototype
11. Software Dev: Testing https://www.ibm.com/think/topics/sdlc • Unit testing • Code review • Security review • Bugs Output: 
 Refined software
12. Software Dev: Deployment https://www.ibm.com/think/topics/sdlc • Staging, and • Production Output: 
 DevOps scripts
13. Software Dev: Maintenance https://www.ibm.com/think/topics/sdlc • Continual review • Fixes • Improvements Output: 
 Regular code updates
14. Scrum Software Development Lifecycle 1. Planning 2. Analysis 3. Design 4. Coding 5. Testing 6. Deployment 7. Maintenance Sprints
15. Image by Annie Spratt. Free to use under Unsplash License. So... When can we ship the simple order edit feature?
16. How much time would that need? 1. Planning - 0.5 weeks 2. Analysis 3. Design 4. Coding 5. Testing 6. Deployment 7. Maintenance
17. How much time would that need? 1. Planning - 0.5 weeks 2. Analysis - 1 week 3. Design 4. Coding 5. Testing 6. Deployment 7. Maintenance
18. How much time would that need? 1. Planning - 0.5 weeks 2. Analysis - 1 week 3. Design - 1 week 4. Coding 5. Testing 6. Deployment 7. Maintenance
19. How much time would that need? 1. Planning - 0.5 weeks 2. Analysis - 1 week 3. Design - 1 week 4. Coding - 2 weeks 5. Testing 6. Deployment 7. Maintenance
20. How much time would that need? 1. Planning - 0.5 weeks 2. Analysis - 1 week 3. Design - 1 week 4. Coding - 2 weeks 5. Testing - 1 week 6. Deployment 7. Maintenance
21. How much time would that need? 1. Planning - 0.5 weeks 2. Analysis - 1 week 3. Design - 1 week 4. Coding - 2 weeks 5. Testing - 1 week 6. Deployment - 0.5 weeks 7. Maintenance
22. How much time would that need? 1. Planning - 0.5 weeks 2. Analysis - 1 week 3. Design - 1 week 4. Coding - 2 weeks 5. Testing - 1 week 6. Deployment - 0.5 weeks 7. Maintenance 6 weeks!
23. Can we do better with AI?
24. Coding AI tiers 3 tiers, according to me Tier 1
25. Coding AI tiers 3 tiers, according to me Tier 2
26. Coding AI tiers 3 tiers, according to me Tier 3
27. Coding AI tiers 3 tiers, according to me Tier 3
28. Coding AI (CLI) 1. Anthropic's Claude Code 2. OpenAI's Codex CLI 3. Google Gemini CLI 4. Qwen CLI 5. AWS Q CLI 6. Cline
29. Team 1. Planning: Product manager 2. Analysis: Product manager 3. Design 4. Coding 5. Testing 6. Deployment 7. Maintenance
30. Team 1. Planning: Product manager 2. Analysis: Product manager 3. Design: Software architect 4. Coding 5. Testing 6. Deployment 7. Maintenance
31. Team 1. Planning: Product manager 2. Analysis: Product manager 3. Design: Software architect 4. Coding: Software engineers 5. Testing: Software testers 6. Deployment: DevOps engineers 7. Maintenance: Software engineers / Product managers
32. Team Agents 1. Planning: Product manager 2. Analysis: Product manager 3. Design: Software architect 4. Coding: Software engineers 5. Testing: Software testers 6. Deployment: DevOps engineers 7. Maintenance: Software engineers / Product managers
33. Demo
34. Demo Plan and analyze Product manager Ask for ideas, goals, scope, etc. Output: Product Requirements Documents (PRD)
35. Demo Architect Software architect Review PRD, ask stakeholder any other questions and define scope for technical implementation. Output: Software Design Document (SDD)
36. Demo Sprint planning Technical product manager / Scrum master Review PRD, review SDD. Take note of timeline, plan sprints, clearly lay out each sprints, define acceptance criteria for each sprints. Output: Clearly planned sprints
37. Demo Implement Software engineer Implement a specific sprint. Output: Software patch in the form of pull request
38. Demo Review Technical lead Review the patch against acceptance criteria. Provide feedback of any errors, but do not fix. If good, tell engineers to proceed to the next sprint. Output: Software review
39. You have heard of Prompt Engineering
40. You have heard of Prompt Engineering Have you heard of...
41. Context Engineering
42. https://blog.langchain.com/the-rise-of-context-engineering/
43. https://artificialanalysis.ai/models
44. https://research.trychroma.com/context-rot
45. https://www.vellum.ai/blog/context-is-king-why-context-engineering-is-the-new-frontier-for-ai-agents
46. Agentic AI Coding Software engineering is now 
 pretty much CONTEXT ENGINEERING
47. Agentic AI Coding • Write specifications, and update them often. • README.md • CLAUDE.md • docs/prd.md (or mvp.md) • docs/sdd.md • docs/sprints.md • docs/*.md
48. Agentic AI Coding • Compact often • Know when to /compact • When to /clear • When to branch out • When to use subagents • Try /context • Remember, too much context, AI loses focus.
49. Agentic AI Coding Prompting is still key Role, Goal, Context
50. Tips
51. AI can help you configure itself • Use Claude Code to generate: • .claude/commands • .claude/agents (or simply /agents) • /init • This generates CLAUDE.md • Review from time to time
52. Start with plan mode.
53. Agent to agent communication • Use GitHub issues. • Or locally, docs/a2a/* or docs/inbox/* • Add to .gitignore • Ask agents to check their inboxes based on roles. • Do not delete, but leave "READ by {role}" after reading it.
54. Keep tasks small • Just like human, you want to keep your pull requests small and on point. • Claude Code can do long-running tasks, but remember about context. • It might lose context.
55. https://x.com/pallavmac
56. Watch out on MCPs • MCP servers are great, but too much of it might overload context window. • Consider if you need MCPs. • I personally am not heavy user of MCPs. • I do see how it can be useful for domain-specific knowledge.
57. https://www.nytimes.com/2025/02/20/business/ai-coding-software-engineers.html
58. If you are a human software engineer • Leverage on AI to learn, fast • Build build build. It is now easier than ever to build. • Build scripts is just a prompt away. • Junior? Leapfrog to senior. • How? Show your projects. Contribute to OSS.
59. Try using different models too • Reviews • Copywriting • Image generation
60. Claude Code • It's amazing. • But... it's not just code. • Try & experiment: • Great agentic AI platform • Give it a server. :) • Give it root access. 🤷
61. Thank you • Code repository is available at • https://github.com/uzyn/nusiss-learningfest-2025
62. Give Us Your Feedback Scan QR Code
63. Thank you chua@uzyn.com U-Zyn Chua https://uzyn.com · LinkedIn: uzynchua