I used to use Cursor as my IDE around a year ago. Not really a coding agent, but definitely a precursor to today's landscape.

Cursor days:

The Good:

The best thing about Cursor was surprisingly not their AI agent, or whatever it was called back then. It was actually their autocomplete model. I still miss it. It was a very intelligent boilerplate writer.

The Bad:

On the coding agent side, the results were alright. Not a huge time save. Their autocomplete was based on Supermaven, which they acquired and integrated. The original product is not in active development anymore.

The Ugly:

Then they dropped a complete predatory overhaul of their pricing model overnight. Same old VC-backed, bait-and-switch. There was a huge exodus of paying customers, including me.

Web Interface only days:

Then followed a time when I would try out the web interfaces of various AIs. I even tried out centralising all of my API keys using LibreChat's web interface. Never really tried out OpenRouter because LibreChat scratched that itch.

Review of that:

It was fine. Not a game changer by any metric, contrary to what tech Twitter would claim in every other post. It saved time, helped brainstorm, etc.

Claude Code initial days:

Claude Code has been impressive to me since day one. Maybe because I did not go into it with the expectation that it would be a fully-autonomous (or even 80% autonomous) tool. I had my expectations tempered, Claude Code exceeded them, and delighted me. Huge time saver.

The initial Claude Code, while impressive, still had a lot of problems. It would ignore instructions and formatting guidelines, it would go around in circles, etc.

So I gave Codex a go.

Codex days:

Codex, as a tool, was alright. Definitely not as good as Claude Code. It would take forever to do one task (upwards of 10 mins usually), and 25% of the time, the results were worse than what Claude Code would do in 25 seconds. Sometimes they were better, but the tradeoff in terms of time did not feel worth it at all.

So back to Claude Code I went. And there I stayed.

Claude Code today:

I have to say it has been very, very impressive. I am loath to praise it overly, because the aforementioned Twitter bros have set about some really, really wrong expectations.

I would never use it with auto-accept edits on. I actually tried out auto-accept-edits on a greenfield project to see it in action. Also, because I wanted to code while I poop 😂 (more on how I achieved that in the next paragraph).

I set up a remote box, with mosh (like a more network tolerant ssh), and downloaded Terminus on my phone. I set up Claude Code on the remote box, and with some more tweaks and enhancements, I was ready to prompt Claude Code from my phone. Surprisingly and disappointingly, the code it produced was pure, pure spaghetti. It would go around in circles, repeat itself, write duplicates, etc.

But I use Claude Code a lot today. To get the maximum, it needs constant human guidance. It's more of a force multiplier than anything else. If the force you put in is 0, the output is 0. If your input is negative, i.e., if you write spaghetti, be prepared to host an Italian family for dinner.

I have also tried out various MCPs and plugins. Not too impressed by them yet. The only one that's decent is pal-mcp. It lets me talk to other AIs (ChatGPT, Gemini, etc) using my own API keys from the terminal. But other than that, I use the vanilla Claude Code.

This wave of coding agents is actually quite, quite opportune and well-timed for small team like ours. The people who will most benefit from this are us, I believe. Why?

- All of us are senior enough / competent enough to know right from wrong
- We are a small team, and our communication overhead is essentially zero

Funny story about a friend of mine who is a team lead/engineering manager, one of her reportees wrote some code that was completely and obviously wrong. And when she asked him about it during code review, the guy's words were "I don't know about this, Claude Code wrote it" 😂. She was pissed! It's fine to write bad code as a junior engineer, but it is not fine to not take responsibility for it!

So, coming back to the topic at hand, this wave is the most beneficial to small and highly talented teams (like ours). Bigger / more dysfunctional companies will still be losing time playing planning poker, and debating whether a ticket is a 5 or 8, and if it's an 8, how we should break it down. Or if a driven engineer wanted to do something like [a game changing feature I came up with and built myself recently], they would still need to pitch it, get it passed, assign it the story points, find place for it in the sprint, get onboard a product manager or even a Techinal Project Manager (phew) so that they can setup a cadence meeting with all the relevant stakeholders, and ... I am tired just writing it out, you get the point.

These coding agents won't help even a little bit with any of that.

Gemini CLI:

Never tried it. Did not hear good reviews, and my first reaction to a Google product (post 2010) is that it is probably going to be low quality. I keep an open mind, though, and Gemini 2.5 pro did pleasantly surprise for a while, but other AIs have already overtaken it. Not to mention, apparently, their authentication is all over the place (as you would expect from Google).

That's it for now! I am very optimistic about this, and I also think we are the primary beneficiaries of this wave. So I expect Formwork (our app) to get better at a faster rate than our competitors (no surprise there).