Practical Agentic AI in Software Engineering. One Year On: Skills, Subagents & Loops

Talk presented at The National University of Singapore Instutite of Systems Science (NUS-ISS) Learning Festival 2026 on August 28, 2026.

Talk was streamed live. Watch it at https://www.youtube.com/watch?v=jjShP3Wpxts
Deck is 1:1 dimension to make space for live demo panel as can be seen from the live-streamed talk above.

Companion code: https://github.com/uzyn/kopi-uncle

Follow my blog at https://uzyn.com

Transcript

1. Practical Agentic AI in Software Engineering One Year On: Skills, Subagents, Loops 1 U-Zyn Chua chua@uzyn.com https://uzyn.com #NUSISSLearningFest
2. Streamed Live 2 • This talk took place on August 28, 2026 at The National University of Singapore. • It was also streamed live at https://www.youtube.com/watch?v=jjShP3Wpxts • Deck is in 1:1 dimension to make space for live demo as seen on YouTube. • Enjoy!
3. About me 3 • Independent engineer • NUS-ISS Adjunct Lecturer • teaching Architecting Agentic AI Solutions • Singapore's first Smart Nation Fellow • Ex-CTO & founder of technology startups • Mentor of tech startups https://uzyn.com chua@uzyn.com LinkedIn: U-Zyn Chua
4. 4 • You've been hearing about Agentic AI all day. • Let's get down to business right away! uzyn/kopi-uncle
5. 5 uzyn/kopi-uncle • docs/ • prd • scrum • skills/
6. Agenda • Software Development Lifecycle • Agentic software development a year ago. [ Slides ] • Workflow • Loop • Skills • Subagents • Evaluation • Graph engineering 6
7. 7 https://datarob.com/essentials-software-development-life-cycle/
8. 8 • Chat, Q&A. Stackoverflow style. • File-based coding. Early-days Cursor. • Full CLI based: Claude Code, Codex. • Multi-CLI & personal agents. https://beyond.addy.ie/
9. Scrum 1.Planning 2.Analysis 3.Design 4.Coding 5.Testing 6.Deployment 7.Maintenance Sprints 9
10. Recap 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 https://github.com/uzyn/nusiss-learningfest-2025/ 10
11. Recap • /agents • Create commands: /plan-and-analyze /implement /review https://github.com/uzyn/nusiss-learningfest-2025/blob/main/prompts.md 11
12. Agent Comms • Agents need to talk. • Engineers <> Reviewers 12
13. Agent Comms How? • Markdown documents • Through human • A2A protocol • GitHub issues (my personal pick) 13
14. Workflow 1. /implement Sprint 1 2. /review Sprint 1 3. /implement fix for review 4. /retro and review overall 5. Rinse and repeat. 🥱 14
15. 15 A lot of waiting a lot of the time!
16. Skills • Skills came a month after my talk last year. • Last year: agents + commands. • Now: SKILL.md 16
17. 17 my-skill/ ├── SKILL.md # required: name, description, instructions ├── scripts/ # optional: code the agent can run ├── references/ # optional: extra docs, loaded only if needed └── assets/ # optional: templates, resources
18. Skills • It's a set of tools for agents to decide which one to call, similar to tool use. • /skill-creator from Anthropic is very useful. 18
19. Evaluation 19 • Know how well your skill or AI system runs. • Come up with test cases. • Evaluate how well they do. • Learn from /skill-creator
20. Loop Let's improve copy, paste, wait... • 2 main loops. • Loop 1: implement-review-loop • Work on a story • Review • Good or change required? • Comms: GitHub PR 20
21. Loop • Loop 2: scrum-loop • If a story is good, merge. • Perform /retro. • Review for any deviation from overall plan. • Pick the next story and start with loop 1. 21
22. Loop • Introduce some exceptions, to break the loop: • Max number of iterations. • Significant drift. • Agent communications are key. • /goal 22
23. Subagents 23 https://pub.towardsai.net/whats-emerging-in-ai-autonomous-multi-agents-and-large-action-agentic-models-lams-7e882a659565
24. Subagents 24 • For most coding agents, they are pretty much built-in. • For Claude Code, to encourage that, you can /effort:
25. Kopi Uncle Puttting it all together View interactive version 25
27. Graph Engineering 27 • You build controlled workflow. • It's not exactly new. Been around for many years.
28. Graph Engineering 28 AutoGen
29. Use your superpowers 29 • Your humanness! • Your creativity! • Review the work from time to time, especially the specs. • AI is not good at design, you are!
30. #NUSISSLearningFest Thank You U-Zyn Chua 30 chua@uzyn.com https://uzyn.com @uzyn #NUSISSLearningFest