The tool calls will be, among other things, something like ReadFile, RipGrep, PatchFile, Shell.
When people talk about the value of different harnesses, they're also implicitly talking about the quality of the system prompt.
The same exact model, when given a different set of tools and a different system prompt, can behave differently.
I think creating your own agent is the Hello World of agentic coding. Instead of Rust, I used D for mine.
(no personal opinions of either, links might be useful)
I think that OpenCode is nice, their CLI version is enjoyable and their desktop/web version is okay:
I also quite like driving OpenCode through something like Kepler / Paseo and tools like that (with those I can still use my Anthropic Condition by Claude Code being treated similarly - as something that gets tasks dispatched to it, while the GUI I see is Kepler / Paseo).
On the desktop side, ZCode was surprisingly usable for something that came out of nowhere (I wasn't aware of it at all before trying out the GLM Coding Plan): https://zcode.z.ai/en
(however, it's much faster to use Plan Mode to build a plan of what it will do, and then execute the plan in Build Mode. you can also have the AI make a script that will be executed deterministically)
And this is coming from someone that's not particularly a big fan of Theo. T3 Code should get more recognition; people aren't just aware of it yet.
I only found this yesterday, and it inspired me to start testing out OpenCode.
That said Claude Code is perfectly fine. I just prefer the integrated experience of using Cursors since I already use VSCode, but I still mostly use Claude Code because of their Max/Fable plan.
this is the integration branch for https://opencode.ai/v2 . it has been for months. it's where the Effect-based refactor has been landing.
I’ve stopped using it completely now.
It works nicely in the browsers on my tablet and phone, too.
On exe.dev you can ask it to customize itself, and it will automatically rebase your customizations when upgrading to a new release.
Anyways, please try mine!
i like to challenge my assumptions and try new tools
that's a very compelling use case, thank you
CC works but for me it felt like increasingly they have zero incentive to make it a great experience. You hear folks like Boris talk about spinning up thousands of agents over night and agents chatting back and forth in GitHub issues and while I think it’s great from figuring out what the future looks like I don’t think it represents the reality of ROI today. So the folks building the tool are so disconnected I am simply not sure it’s a great experience anymore.
Is that the main concern though, cost?
That being said, I had to nope out of a similar thing from GPT 5.6 today, so it appears to be a US frontier lab issue. Claude is particularly bad though, as it produces far too much code even when I tell it not to, unlike GPT (and Kimi) which at least listen to me a little better.
More generally, I want a usable human review experience, and Claude code doesn't deliver that for me.
IMO part of it is that the underlying LLMs have gotten better enough that harnesses feel better even if they haven’t changed. I have a toy harness that barely implements the features you’d expect and it works surprisingly well. Like there’s literally nothing clever, it calls tools and that’s about it, and it still mostly does the right thing.
Edit: lol, I don't think ACP is even actively developed anymore. It seems to have been merged into another seemingly pointless standard with an even worse name, A2A. [0]
What’s the counter argument? pi and ohmypi are pretty fantastic. Of course like all developer tools it depends how you do your work but I am not sure what you are trying to achieve in your comment.
Cybersecurity capability might be nerfed
I feel like there should be some mechanism to prove you own the code/app/site/whatever and it will remove the guardrails from the LLMs allowing them to find and fix these vulnerabilities.
You can't guarantee everyone else will use a neutered model.
I've joined TAC, but still have to dance around it.
In this case, you can put whatever you want between the harness you're running (or modify the harness itself), and essentially "lie" to the model. Any verification technique would be fairly trivial to bypass, while you continue to run the harness locally.
Have personally tested this with Opus and Sol and it works.
Classifiers are tricky though. Here's where open weights will win.
Anthropic was stingy as hell with its Fable and cybersecurity nonsense, switched to OpenAI which is much better but still not enough. I'm tempted to switch again...
If it won't attack my stuff, it won't help me build my stuff to be secure.
What do you mean with this? Honest question!