Dariusz Walat

AI Coding Agents in Practice

Articles documenting what actually happens when you build production-ready software using AI coding agents as the primary development method — not as occasional assistance, not for boilerplate, but for all the code.

Each piece is a snapshot from a specific point in time. AI technology moves weekly: what’s true about Claude Code’s behavior today may not be true next month. Patterns named here may become obsolete as models improve or new tools emerge. The date on each piece matters.

The public narrative about AI coding agents is incomplete. Productivity gains are real. So are the operational realities — agents generating plausible documentation for code they didn’t write, multi-agent systems compounding failures, review burden overwhelming generation speed gains, nondeterminism making reliable workflows nearly impossible. These pieces document the second part.

Published

AI Coding Agents Are Plausible Bullshit Generators October 18, 2025

When agents generate convincing documentation for functionality they never implemented. Why you must verify everything yourself, and what this means for production systems.

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.

On reading these

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

These are observations from one engineer building personal projects with specific tools at specific points in time. Your context, your tools, your problems may differ. If you’re building production systems with AI coding agents right now, you’ll recognize patterns here — they’re not specific to my setup; they’re characteristics of current AI technology. Use these pieces as starting points 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.

Articles in this series:

  • Why This Series Can Never Be Finished
    When Your Research Subject Evolves Faster Than You Can Document It
    I was writing an article about AI agent nondeterminism. Hours later, I realized parts of it were already outdated. This isn't a bug - it's proof of why documenting rapidly-evolving technology requires a fundamentally different approach.
  • AI Coding Agents Are Plausible Bullshit Generators
    How Verification Failures Become Production Disasters
    When agents generate convincing documentation for functionality they never implemented.