Curious to see how adoption among cli agents will go (it’s nice to see Gemini cli already in).
The level of competition in the LLM and coding assistant market is always nice to see, and this only helps to make costs of switching between offerings even smaller.
You need to define an interface for Ai to click buttons? Or to create keyboard macros that simulate clicking buttons? We are all doomed!
I welcome them openly collaborating.
I ask AI to write a feature (or fix a bug, or do a refactoring) and then I read the commit. If the commit is not to my liking, I "git reset --hard", improve my prompt and ask the AI to do the task again.
I call this "prompt coding":
https://www.gibney.org/prompt_coding
This way, there is no interaction between my coding environment and the AI at all. Just like working with a human developer does not involve them doing anything in my editor.
This is the equivalent of LSP but for coding agents. So any editor does not have to rebuild an interface to support each and every new one.
Very big doubt. AI can help for a few very specific tasks, but the hallucinations still happen, and making things up (especially APIs) is unacceptable.
The new models are much better at reading the codebase first, and sticking to "use the APIs / libraries already included". Also, for new libraries there's context7 that brings in up-to-date docs. Again, newer models know how to use it (even gpt5-mini works fine with it).
The real issues I'm struggling with are more subtle, like unnecessary code duplication, code that seems useful but is never called, doing the right work but in the wrong place, security issues, performance issues, not implementing the prompt correctly when it's not straight forward, implementing the prompt verbatim when a closer inspection of the libraries and technologies used reveals a much better way, etc. Mostly things you will catch in code review if you really pay attention. But whether that's faster than doing the task yourself greatly depends on the task at hand
I've had much more success with things under 20k LOC but that isn't the stuff that I really need any assistance with.
Currently, my environment uses Claude Code on bare metal, and my application runs in a container, and the agent can do "docker compose exec backend" without any restrictions (YOLO).
My biggest obstacles to adopting workflows with git worktree are the need to share the database engine (local resource constrains) and the initial migration time. Offloading to cloud might be interesting for that.
> with popular languages
Don't know, don't care. I write C++ code and that's all I need. JS and React can die a painful death for all I care as they have injected the worst practices across all the CS field. As for Python, I don't need help with that thanks to uv, but that's another story.
Rather than rushing to create standards, I think the focus should be on building practical implementations, AI-centric UI components that actually help developers design more AI-friendly interfaces. Once the space matures and stabilizes, that’s when standardization will make more sense. Right now, it feels too early.
The famous "It compiles on my machine." Is that where engineering is going? Spending $billions to get the same result as the laziest developer ever?
If your proposed spec is full of "SHALL", "MUST", and "MAY", I'm going to dock you ten points of credibility from the outset. It's a signal you've set out to imitate the vibes of seminal RFC specs without independently considering the substance.
And don't respond with a childish "skill issue lol" like it's Twitter. What specific skill do you think people are lacking?
That obviously does not mean that it's good software. That's why the rest of my comment exists. But "AI is hallucinating libraries/APIs" is something that can be trivially solved with good software practices from the 00s, and that the AI can resolve by itself using those techniques. It's annoying for autocomplete AI, but for agents it's a non-issue
Are people who assert that JSON-RPC is "heavy" writing code we all rely on?
Most of the time I don't bother reprompting it either. If it doesn't understand the first time then I'm better off making the change myself, rather than sink time into a cycle of reprompting and rejection.
On the other hand, if it understands what I want the first time I have more confidence that it's on the same wavelength as me.
https://lfaidata.foundation/communityblog/2025/08/29/acp-joi...
It's like the Jupyter protocol, from which people do derive utility.
The purpose of an ide is to pull together the tools a developer needs to develop software… first, reading/navigating and writing code, next running and debugging code, then (potentially) a variety of other tools: profilers, doc browsers, etc… all into a unified UI.
But coding agents seem to already be able to use the command line and mcp quite well to do this.
So why mediate using these tools through an IDE (over a new protocol) rather than just using the tools directly through the command line or mcp? It’s two extra levels of indirection, so there needs to be a really good reason to do it.
There may very well be some actual problem this solves, but I don’t know what it is.
Or is A2A like USB, where it acts as both a registry of, and “standardized standardization process” for, suites of concrete message types for each use-case?
Like, yeah, when a "client" drives an "agent", that's no different than what any generic "agent" would be doing to drive an "agent"; an IDE or what-have-you can just act as the "parent agent" in that context.
But when an "agent" is driving a "client", that's all about the "agent" understanding that the "client" isn't just some generic token-driven inference process, but an actual bundle of algorithms that does certain concrete things, and has to be spoken to a certain way to get it to do those concrete things.
I had assumed that IBM's older ACP was in large part concerned with formalizing that side of interoperation. Am I wrong?
One great reason is to avoid M*N problem: https://matklad.github.io/2022/04/25/why-lsp.html
Feels like a lot of mindshare has shifted towards VSCode, so that's where the tooling investment goes. I don't want to be forced off of subl because new tools stop supporting it - it's a great editor, and it's not sponsored by a massive company.
there is a learning curve, it reminds me of learning to use Google a long time ago
As you suggest, I've had a moderately successful time trying to get AI to write its own Sublime Text plugins so our favorite editor doesn't get left behind, so might be cool to try with this too?
The IDE is in-between because it asks the user for confirmation before doing anything, if that's what you are asking. It's not adding any indirection or something, this is just how coding agents talk to the IDE already, all this does is standardize the language they speak which would be a win for everybody.
https://github.com/zed-industries/zed/blob/main/crates%2Fage...
https://www.npmjs.com/package/@zed-industries/claude-code-ac...
Even though that doesn't seem to be mentioned on the site yet.
This is the top rated discussion on Github though: https://github.com/zed-industries/zed/discussions/26770
Helix has been my favorite editor ever since I started using it, but the lack of integrated AI features has been becoming more of a deal-breaker for me lately. Though, I can see why the maintainers are hesitant to build for it, because there hasn't been a standardized protocol. I hope ACP changes that.
Agents use multiple models, can interact with the environment, and take many steps. You can get them to reflect on what they have done and what they need to do to continue, without intervention. One of the more important things they can do is understand their environment, the libraries and versions in use, fetch or read the docs, and then base their edits on those. Much of the hallucinating SDKs can be removed with this, and with running compile to validate, they get even better.
Models typically operate in a turn-by-turn basis with only the context and messages the user provides.
I have extreme intolerance to boredom. I can't do the same job twice. Some people don't care.
This pain has caused me to become incredibly effective with LLMs because I'm always looking for an easier way to do anything.
If you keep hammering away at a problem - i.e. how to code with LLMs - you tend to become dramatically better at other people who don't do that.
Yes, Zed's startup time is way better, and it definitely feels snappier. But when I work on code, I need to jump around the code. I want the debugger to show me the current line in the file, annotating the value of every variable right in the source. I want to jump to property usages. I want to automatically rename a function everywhere intelligently, not using search-replace. I want to get completion for my database tables, and browse them from the IDE. And lots more.