What This Series Is

For the past several months, I’ve been building personal projects using AI coding agents as my primary development method. Not experimenting with AI on the side. Not using it occasionally for boilerplate. AI agents generate all my code.

This series documents what actually happens when you build production-ready software this way.

Not theory. Not speculation. Not “best practices” copied from other articles. Operational patterns from building real software with AI agents right now.

What This Series Is Not

This is not finished research.

AI technology changes weekly. What’s true today about Claude Code’s behavior may not be true next month. Patterns I document now may become obsolete as models improve or new tools emerge.

This is not definitive guidance.

These are observations from one engineer building personal projects with specific tools at a specific point in time. Your context, your tools, your problems may differ.

This is not advocacy or criticism.

I’m not arguing AI agents are good or bad, ready or not ready, the future or a fad. I’m documenting what works, what fails, and why, as of October 2025.

Why This Series Exists

The public narrative about AI coding agents is incomplete:

What you hear:

• “AI makes developers 10x more productive!”
• “Just describe what you want and ship it!”
• “The future of coding is here!”

What you don’t hear:

• Agents generate plausible documentation for code they didn’t write
• Multi-agent systems compound failures exponentially
• Review burden can overwhelm generation speed gains
• Nondeterminism makes reliable workflows nearly impossible

Both narratives are true. The productivity gains are real. The operational challenges are also real.

This series exists because nobody is documenting the second part publicly.

Why It’s A Living Series

AI technology evolves faster than traditional software tools.

A framework released 5 years ago works mostly the same today. An AI model released 3 months ago has been superseded twice.

Writing about AI agents as if findings are permanent would be dishonest. They’re not. They’re snapshots of current technology state.

This series will grow and change:

• New essays as I discover new patterns
• Updates to existing essays as technology changes
• Corrections when my understanding improves
• Additions as tools evolve

The date on each essay matters. What’s true today may not be true in weeks or months. Technology evolves faster than documentation.

Published Articles

Why This Series Can Never Be Finished October 19, 2025

When your research subject evolves faster than you can document it. Why documenting rapidly-evolving AI technology requires a fundamentally different approach than traditional software documentation.

Additional articles are being finalized for publication.

Why I’m Publishing This

Professional transparency: I’m documenting this work as personal research based on experiences from personal projects, before similar patterns become part of any company’s practices. This establishes the direction of knowledge transfer.

Community value: Every team adopting AI agents will hit these same problems. Publishing my findings helps others skip expensive lessons I already learned.

Personal positioning: Managing AI-driven workflows is becoming a valuable engineering skill. This series demonstrates understanding of the operational realities, not just the marketing promises.

Continuous learning: Writing forces clear thinking. Publishing creates feedback loops. Both make me better at working with AI agents.

The Honest Disclaimer

Personal research based on experiences from personal projects, conducted independently in my own time, unrelated to employment.

Your mileage will vary. Your tools may behave differently. Your problems may not match mine.

But if you’re building production systems with AI coding agents right now, you’ll recognize these patterns. Because they’re not specific to my setup - they’re characteristics of current AI technology.

Use this series as a starting point for your own research, not as definitive answers.

The technology is too new, too fast-changing, and too context-dependent for definitive answers to exist yet.