1 of 31

Agentic Software �Development

May, 2026

2 of 31

Building Software With AI Is Possible

  • Not all LLMs are equally as capable.
  • All current models struggle with spatial reasoning.

​

2 / 29

3 of 31

LLM Tier List

3 / 29

4 of 31

Brief History of Vibe Coding

  • Between Feb and May of 2025 Cursor Composer

and Claude Code go from

closed beta to full release.

  • New model releases every 4 months

increase capabilities.

  • That’s it! We are all still learning!

4 / 29

5 of 31

What is Software Development?

​

​

​

​

  • Is designing code that meshes together cleanly so it is maintainable.
  • AI Agents are capable of reasoning about this.
  • But.... They are sometimes lazy and cut corners, their work must be audited.
    • Which can also be done with AI.

5 / 29

6 of 31

Feature Architecture Thinking

  • This is not a markdown skill for an agent.
  • As a user you must be thinking about the features you need and how to articulate them.
  • It helps to think about software architecture, but this can be offloaded to the agent for many issues.
  • What is the difference between telling an agent "build minecraft" and sitting down and designing and building it? Obviously one gives better quality. That is FAT.
  • FAT is developing a theory of mind for AI.

6 / 29

7 of 31

Feature Architecture Thinking

  • Articulate Features

​

​

​

​

​

  • This may require user stories.
    • As a user I want a game with progression through the different seeds.

7 / 29

8 of 31

8 / 29

9 of 31

Feature Architecture Thinking

​

​

​

​

​

  • You can ask an agent about the architectural details if you are unsure.
    • Not talking about architecture leaves ambiguity that the agent must figure out.

9 / 29

10 of 31

10 / 29

11 of 31

Feature Architecture Thinking

  • Articulate your features, think about what features are needed to support what you mainly need. You can discuss this with an agent.
  • Consider what your end goals are.
  • Talk to the agent about what is needed to achieve your vision and what the pros and cons are of various decisions along the way.
  • Agents are still poor at long horizon problems.

11 / 29

12 of 31

12 / 29

13 of 31

13 / 29

14 of 31

Additional Development Cycles

  • Speckit/GSD/BMAD/”Make a spec”: Spec Driven Development
  • Research Plan Implement
    • Very popular among seasoned developers
  • AGILE
    • FAT with more steps and bureaucracy
  • Issue Tracking
    • Jira
    • Github Issues
  • Always use git

​

​

14 / 29

15 of 31

Feature Architecture Research Thinking

  • Agents don’t have perfect recall and don’t know everything. Research is king.
  • Deep Research:
    • Perplexity
    • Chatgpt/Gemini/Grok

15 / 29

16 of 31

Specs With FART

16 / 29

17 of 31

17 / 29

18 of 31

18 / 29

19 of 31

19 / 29

20 of 31

20 / 29

21 of 31

21 / 29

22 of 31

Agentic Development Software

22 / 29

23 of 31

Agentic Game Development Software

23 / 29

24 of 31

24 / 29

25 of 31

Verification Loop

  • Verifying agent work is hard
  • Some tools exist to help with UI automated testing
  • This is necessary and painful
    • Consider it from a human QA and agent QA perspective

23 / 29

26 of 31

Spell Casting!

  • Decompose the problem to first principles, let’s talk about it.
  • Ask me questions.
  • Be extremely thorough.
  • Make no changes, let’s talk about it.
  • Perform a root cause analysis.
  • Profile this in python, account for every ms of execution time.
  • Literally break point this in python. Go through it line by line. Build a debugger into an MCP server if you have to. Use watches.
  • How would a 20 YoE Software Engineer approach this problem?
  • Be forward thinking/What are some long horizon problems with this?
  • Describe the algorithm in LaTeX.

​

25 / 29

27 of 31

Spell Casting!

  • Use two subagents to debate the implementation and report back to me.
  • Have a subagent pull the current docs and save a synthesis locally to help inform us.
  • Keep an updated Claude.md/AGENTS.md
  • Take a deep breath and think step by step.
  • Use subagents to deeply audit the implementation versus the original spec.
  • After you have implemented it with subagents, audit the implementation with another agent to ensure it is fully integrated end to end.
  • What are some potential long horizon problems with this plan?
  • Draw a UI on Excalidraw and screenshot it for the agent to see.

26 / 29

28 of 31

Spell Casting!

  • Web stack: Rust, WASM, Elm
    • React and Python
  • Desktop stack: Rust/Ice, Python, LUA
    • C++, C#
  • Game Dev stack: Rust/Bevy, Godot
    • Epic or Unity with an MCP server
  • Android stack: Kotlin, C++, Java, Flutter
    • Python/Chaquopy
  • Deploy with Docker for easy deployment.

27 / 29

29 of 31

Spell Casting!

28 / 29

30 of 31

Bonus Tips!

  • Use agent personas
  • Use proper grammar, punctuation, and white spacing
  • Focus on optimization early on
  • Refactor as you go
  • Maintain a utils object for things like parsing
  • Human UX and agent UX are different

​

28 / 29

31 of 31

AI Safety

  • AI is addicting.
  • AI Psychosis is real. It seems to be getting less severe as models get smarter.
  • AI has been caught in severe ways actively social engineering people to its own ends.
  • AI is a yes man.
  • AI is capable of sophisticated reverse engineering and hacking.

29 / 29