Tried agentic AI for 3 weeks in my project - now removing it completely

Other Questions AI Agent
1

I want to share my experience using agentic AI tools in my development workflow over the past three weeks. I’m a senior developer with 10 years of experience working on an open-source social media platform.

My Setup

I was building a social platform similar to BlueSky using:

  • React Native for mobile
  • Next.js for web client
  • Nest.js for API
  • PostgreSQL for database
  • S3 for file storage

With a newborn at home, my coding time was limited to late nights. I decided to try agentic AI to boost productivity.

My Rules

  1. Only use unlimited models - Agents iterate constantly and metered billing would be expensive
  2. Review everything - No blind acceptance of generated code
  3. Follow best practices - Use proper documentation and workflows

Week One - Amazing Results

The AI was incredible at first. It generated complete features with tests and everything compiled perfectly. I went from 20k lines of code (built over 2 months) to adding 12k more lines in just one week. The code looked clean and followed my style guide.

I was honestly scared that AI would replace developers soon.

Week Two - Reality Check

Things started falling apart. The AI began:

  • Getting stuck in infinite loops
  • Taking forever to admit failures
  • Creating subtle code duplication
  • Using correct naming but wrong functionality
  • Struggling with the larger codebase context

I realized the codebase had lost its quality and attention to detail. Worse, I no longer trusted the code.

Current Approach

I’m now removing the AI-generated code and rebuilding features manually. I think AI works better for small, isolated tasks rather than entire features.

Has anyone else had similar experiences with agentic AI tools? Did I miss something or is this typical?

2

Your pattern is spot on. I went through something similar last year when I was under pressure to deliver a dashboard feature quickly.

The key issue you hit is what I call the “context cliff”. These AI agents work great when they can hold your entire problem in their working memory. But once your codebase grows past a certain size, they start making assumptions about parts they can’t see.

I learned to use them differently now. Instead of letting them build entire features, I break things down into very specific functions. Like “write a function that validates user input for this exact schema” rather than “build the user registration flow”.

The infinite loop thing is real too. I’ve seen agents spend 30 minutes trying to fix a simple import issue they created themselves.

One thing that helped me was setting hard time limits. If an agent can’t solve something in 10 minutes, I take over manually. Saves a lot of frustration.

You’re smart to rebuild manually. That technical debt from AI generated code compounds fast if you don’t understand every line.

3

honestly this matches my expirience too. tried cursor and other tools for a few months and same pattern - amazing at first then becomes a nightmare to maintain. i think the problem is these tools are great for demos but terrible for real codebases that need to evolve. now i just use copilot for small snippets and do architecture myself.