My main problem with the current state of the editors is that either they're are bloated i.e. VS Code or they've drift out of being editors i.e. Claude Code desktop or the new agents panel that Cursor/Antigravity and others have introduced
None of the existing ones (that I've used at least) are good editors designed to work with agents.
For a moment I thought it was some open-source LLM trained on legal. It's not, it's a web app wrapping major LLM providers and streamlining legal workflows, uploading documents, and having the LLM providers interact with them.
Harvey made it a point to FT ChatGPT models for a year or so but they were struggling to keep up with the pace of new model deployments and quit. They never went as far as Cursor AFAIK which produced its own routers/"composer" models.
In "Thinking—Fast, Slow, and Artificial: How AI is Reshaping Human Reasoning and the Rise of Cognitive Surrender" [1], a third level of the human thinking process was introduced where argues that besides internal intuition (level 1) and deliberation (level 2), there's a third one which is external and is the well-known to everyone, AI.
Humans now can choose to either completely surrender to that external intelligence, called cognitive surrender or choose to just offload tasks and use it alongside their own reasoning.
It's up to us how to use it. I've recently joined a new company and I had to explore their massive codebase. Even my manager suggested that I should use an agent to help me with the onboarding but I went straight to the code and whenever I had a new task, I would explore the code to understand what is happening and then instruct the agent to do any tasks I thought should be done.
There are moments when I completely hand off to the agent without looking any diffs, but my main focus is understanding first and then handing off.
I do like Antigravity a lot and I use it on a daily basis. Are you referring to that or something else? I do find their marketing weird though and seems like their internal teams are not yet fully aligned. (Jules, Antigravity, Gemini CLI, ...)
I've come across only a few of those. I'm not familiar with Antigravity. That's what I mean with poor marketing. I've tried out AI Studio, which is alright but not quite Codex level good as far as I can see.
Honestly this smells a bit like the situation they had a few years ago with a lot of different teams in Google inventing new chat and video call tools. All of which then flopped and got discontinued. Google exporting their own internal chaos. How many coding tools can you have in one company?!
OpenAI and Anthropic seem more focused currently. I mostly focus on OpenAI's codex because I don't want to maintain two subscriptions. Switching tools is a bit disruptive and not something I want to do every few days/weeks.
Wouldn’t it make a lot of sense to have a single subscription that gives access to all coding agents, with the ability to switch between them based on preference or task? Constantly juggling tools and subscriptions is pretty disruptive.
20 euro per month is pretty hard to beat for Chat GPT Plus (which includes Codex). Right now a lot of this stuff in this space is highly experimental and constantly changing. I've been using codex web since before the summer, Codex cli for the last few months and Codex desktop app for the last few days. Before that, I was copy pasting blobs of code from Chat GPT and only looking at single code files.
The whole agentic coding revolution didn't really start moving until Claude Code rolled out almost a year ago now. Anthropic deserves kudos for that. Initially it was limited by small context sizes and wasn't that effective on large code bases. So, I stuck with ChatGPT and OpenAI; mainly because Claude for Desktop was a bit underwhelming and I felt OpenAI had their shit together a bit more in terms of UX, which I think matters at least as much as model quality for being able to effectively use AI. Arguably, since about the summer, Codex and Claude Code are well matched in terms of features/capabilities. Some prefer one or the other or use both. Coded had a reputation of being maybe slightly better with larger code bases. I don't think that is valid anymore as of the last few model releases.
Since about the gpt 5 generation of models released around last summer, I'm able to work on whole git repositories with codex. Our backend is about five years old; about 85K lines of code. Not huge but big enough that there was no way in hell LLMs were able to make sense of it before that. I only did the first big pull requests with codex on this backend in the last two months. This stuff is still very new.
I'm sure that with a lot of tool juggling and experimentation I could have gotten there a few weeks/months earlier But not much more. And I don't actually have time to be constantly fiddling with tools and trying out a lot of stuff. I don't need to be the first to try everything out.
I'm trying to do a deep dive on RAG/LLM based apps since I didn't had the chance yet.
I'm building a chat-assistant that will be able to discuss with you and find more about what's your current work status and what you're looking for. Then it will suggest best suited roles for you from the HN Who's Hiring threads.
It's still very early, I've managed to index the latest thread and there's a CLI chat tool to discuss with the LLM.
It's been great because I learned a ton already, about LLM deployments, RAG evaluation, prompt engineering, LLM internals, etc.
I'm thinking that GDP is best used for longer term periods and not for a goverments' short term term, 4-10 years. For example, a goverement could be recklessly spending money that could have an effect on GDP but that doesn't show the whole picture.
Regardless, GDP is always useful so def one of the right metrics for economic performance in general.
None of the existing ones (that I've used at least) are good editors designed to work with agents.