How many coding agents are you using for the same project?

A developer with roughly 15 years of programming experience describes switching between three coding agents on a single project after abandoning an 11-year WebStorm subscription within 30 minutes of trying Cursor. The engineer reports that agent-assisted work compressed multi-day feature implementations into about 15 minutes, but that shipped features often felt incomplete or broken despite the specific prompt being satisfied.

Probably my first post ever...anywhere. In the last few months, I've had to switch between 3 coding agents for the same project just to barely get it working consistently and to my design more on that below . I remember being excited just to have a machine talk back to me :D fun times . We've come a long way since then. We now have agents generating code, running CI/CD, triaging production defects, and even writing posts—well, maybe not this one : . So why am I writing this? Probably just exhausted with AI, or maybe because I'd like to connect with people and see if I'm alone in my frustrations, or if fellow engineers are experiencing the exact same thing. I had been using WebStorm for 11 years up until 2024. It was quite simply the best, considering I had tried Eclipse before that and well, Notepad before that . I felt like nothing could beat WebStorm. It made me productive, exceptionally fast at writing code, building projects, and enforcing conventions. No complaints. Everything was going great Then one day a friend of mine said, "Dude, have you seen Cursor? " I brushed it off like it was just another gimmicky fork of VS Code by the way, I didn't even like VS Code that much at the time . A week later, I started feeling slower compared to colleagues using Cursor. Human curiosity took over, so I downloaded Cursor and gave it a shot—while remaining completely convinced that my WebStorm config was far superior hah . I cancelled my 11-year WebStorm subscription within 30 minutes of using Cursor. The boost was unreal to me I used to slog for 6–7 hours straight trying to implement API authentication with edge caching and hoping it worked my code never runs error-free the first time, to be honest . In Cursor, I watched it not only finish the implementation end-to-end, but write the unit tests, run the build, pass lint and formatting checks, AND update the docs—in under 20 minutes Probably the first time I had an out-of-body experience :D How could I not trust it? I felt like a 3rd grader stumbling upon a book on advanced calculus. I remember having a headache, excitement, fear, and confusion all at the same time. First 1 or 2 weeks I was just exploring this new feeling, it was fun. I finished features in days that could have taken weeks. I figured out how to bring my WebStorm config setup into cursor, setup everything to work the same way but better best of both worlds . Thinking if I get everything right, set everything properly, configure everything correctly I would never have to touch code again All I'd have to do is tell the machine to do it for me and have spare time to do something else. Watch a movie maybe? Go to the gym? Take a breather for once? Improve my daily water intake :D And it really felt like this is it I was building things at the speed of thought. Pretty soon waiting for the agent to finish "Thinking...", "Analyzing Code", "Executing..." wasn't enough anymore. I wanted more, I wanted to do more things in parallel... From ~15 years of 9-10 hours a day of careful programming, DDD - to 15 mins of end-to-end feature completion within days and now I wanted to do more in those 15 minutes With "spare" time on my hands, I really wanted to learn and understand how this works I had no background in AI engineering or Machine Learning . Read through LLM architecture, transformers; fascinated by the fact that such a simple concept of "attention" could have such a massive impact on model output and the field of AI from fiction to reality . But, just like in statistical maths: every metric or statistic has outliers; every "trend" projection is formulated with control variables. You begin to notice things We have a natural ability to adapt, given enough time. And, as my "vibe coded" features started rolling out in my projects, I noticed something strange: even though the specific feature I prompted the coding agent to complete was done, it felt incomplete or broken only when an actual user used it. Something was up here And as I normally do, I blamed myself—maybe my prompt was not perfect Probably I didn't configure it correctly I guess my code was already broken and the agent did the best it could Did it? The more I dug into it, the more I analyzed the written code, I could feel the emptiness of the implementation. Sure, the individual functions were fine, there were unit tests, there was perfect documentation with all the technical jargon possible. But in the end, what felt lost was the practicality, the domain understanding, the edge cases, the nuances of how the features fit into the larger product. The soul of rational flow. As you may know by now, I started with Cursor - I could create different sessions, one for code review, one for development. The reason for 2 or more different sessions was to avoid context bias and ability to switch between models in different sessions. I would also use completely different prompt styles in each session just to create that adversarial net. prompt Help me build features while following rules and guidelines that I've setup. If you have any questions or sense ambiguity, discuss before jumping into code. Sometimes I even used skills like "grillme", "frontend", "architecture" and others. prompt I'm setting up this session to review work/implementation that has been done. Review the files changed and ensure the following every time I ping you: - Make sure the implementation is structurally sound. - Ensure all tests include adversarial functional tests. - Review changes for design and end to end flow integrity. - Review implementation quality. And the way I would work is - go to "develop" session, discuss feature - go few rounds - use plan mode - implement the feature - see tests green. then: go to "review" session, discuss implementation - ask it to review - review summary - ... Go few rounds back n forth between those sessions till I "felt" its stable now. Then, finally, commit n push. Before I knew it, days would pass and I'd still be juggling between sessions cause I couldn't trust what was being generated. I was babysitting AI which was supposed to be building my project with a single prompt . At first, I thought its not you, its me Maybe I didn't configure the rules properly, maybe the prompt was wrong, maybe I didn't use the right skill framework or the right wording. After trying all possible options I remember having a headache, excitement, fear, and confusion all at the same time. In came - Claude Code, when I heard of it around November 2025. Cancelled Cursor, subscribed to claude and started riffing with Sonnet 4 and then couple of months back Opus & Fable 5 and at the same time I was using VS Code Copilot with Opus 4.8 where I work. My struggles with cursor, the constant back n forth, the design n redesign n replanning - Gone everything was done by Claude, I felt like a spectator again, just witnessing a magic show. And then, I started noticing the pattern - Claude was also exceptionally dense in coding, implementing functionality and extremely long summarization of the things done after the turn finished. The Problem? the claude code harness started showing a pattern of - and at the end it would always say "Few things I left intentionally open" not matter if its a simple change or a multi-module refactor The pain - now I end up being responsible for reviewing every line end to end cause I can't trust what "it left" and it didn't AI is extremely advanced nlp, efficiently compressed knowledge vectors and weights; but in the end its still statistical probability dialed up to a million times. Is it worth it? Not sure, cause that's subjective. I love using it to get over my thinking block at times. It does help me refactor code faster and with better intent than a traditional IDE, but I still would prefer being in control of the design, the careful, meticulous architecture building and code segmentation, using AST tools for code evaluations and sanity checks. AI isn't good or bad in my opinion, its just the next system enhancement to speed up repetitive tasks and allow focusing on design thinking and evaluation than actually "typing" the code. Is it accurate? Miles away from it but that's what I thought when GPT-2 was released and it could barely talk