Lately, I've been experimenting with Cursor's AI coding agent and had a pretty fascinating "vibe coding" session — part exploration, part debugging, part babysitting a well-meaning intern.

Here's what I learned:

1. Agents Still Get Stuck in Loops

I hit a strange bug: when I asked Cursor to only apply a subset of the changes it proposed, it got confused and entered a 3–4 loop cycle where it kept proposing the same edits over and over. Eventually, it got it right, but this was a reminder that most agents today struggle with partial, constrained edits — especially when their internal state drifts from the file state.

2. Great at Explaining, If You're Specific

When it came to debugging or explaining data flow, Cursor was surprisingly solid — as long as I explained my ask clearly, step by step. If you treat the agent like a junior developer and spell things out, you get good results. Be vague, and it will wander off.

3. Still Prone to Brittle Logic

One big miss: instead of trusting a language model's memory of the conversation, it hardcoded brittle rules to react to specific phrases in the conversation history. That's the kind of logic that falls apart fast in production. It reminded me how easily these tools can regress into imperative hacks instead of building robust abstractions.

TL;DR:
AI coding agents are powerful but still quirky. They loop, they misstep, and they need handholding. But with the right prompting and mindset — think junior dev with superpowers — they can be incredibly useful.