Weird to see Cary mentioned as the setting

I honestly don’t understand comments like this because in my work, this would be a disaster. And before I get the comments about my harness/skills etc, I’ve tried many tools and harnesses and skills and all that earnestly and in good faith. I find use in it for doing the grunt typing labor, but letting it loose in ways described above have only ended in spending much more time cleaning it up than if I just did the work myself.

I hate to sound pretentious, but I wonder if it’s a difference in complexity of work and problems being solved.

I just reviewed this thread, and thanks for an opening to say something I had realized I missed.

> I hate to sound pretentious, but I wonder if it’s a difference in complexity of work and problems being solved.

It is about complexity, at least for me. I am working on b2b SaaS.

There are times where even using LLM assistance, I spend weeks or months working on a tough problem.

However, the <show product get feedback> loop is now nearly entire automated, when it does not involve some actually complex problem, which are most of the meetings.

I still am mind blown at how bad CC is as software. It’s just not that hard of a problem. I get that harnesses aren’t trivial, but they’re not insane either. And the fact that it’s running on a JS runtime (that they bought!) is also crazy. Why not Go/BubbleTea? Why not literally anything native? It makes no sense

I don’t want to be a jackass, but it’s hard to take anything Boris says seriously when he’s headed up such weird software. And it’s not like resourcing or money is an issue for them. If Claude was so damn good, why does CC suck?

I'll be honest, I don't actually understand what people mean when they say Claude Code is bad software.

Seems pretty good to me. Presumably this is about the TUI version?

I'm surprised you think Claude Code is good software. I find it so hard to use because it is fundamentally constrained as a TUI. It is buggy, clunky and slow. It is sooo slow.

One example of what is particularly bad with it: tool calls and progress. Codex UI does it so much better. The way in which CC waits for a task to complete etc is really poorly done compared to Codex.

Another thing that's missing: no way to continue a side chat in Claude Code - this is easy in Codex with /side and it is super usefu.

Claude Code has /btw which I think is the equivalent of /side in Codex.

I don't think Claude Code is perfect software, but I don't think fact that Codex has a slightly nicer implementation of certain patterns makes Claude Code bad software.

> Claude Code has /btw which I think is the equivalent of /side in Codex.

it tries but /btw can't be invoked at any time. it is blocked until the reasoning is done. you also can't continue the chat with it

you also can't have tool calls inside it

Harnesses try to solve a wide range of complex, open-ended UX problems, so there isn't a perfect one.

CC might be the best one I've used if I give each major UX aspect a rating 1-5 and then average them. For example, it has a decent subagent viewer.

Meanwhile, Codex doesn't even have one, and its subagent tool call is so buggy that the parent agent sometimes doesn't even know why the child died.

Being a TUI is very limiting, yes, though that limitation isn't the harness' fault.

I've only used the Claude TUI, but it is extremely sluggish. It takes 1-2 seconds minimum to launch on a MBP M4 with 48 GB of memory. I have also encountered plenty of display-related bugs that you can find documented all over the internet. If you had to group this into a software quality bucket, it certainly would not fall into "good". Maybe "mid".

Do you really think that having this many issues is justifiable/ok?

https://github.com/anthropics/claude-code/issues

Those numbers don't really mean anything. Claude Code has millions of users, and that issue forum is the most obvious place for them to ask questions or request features.

If there were 11,000 open and confirmed bugs then yeah, that would mean the software is bad.

No one knows how many bugs claude has because Anthropic auto-closes Issues (or at least used to) if there isn't someone constantly pinging the issue every two weeks. I've contributed to several issues that were confirmed by several other people, which were they were auto-closed after people gave up confirming the problem without any response from Anthropic. They seemingly don't care if it isn't on fire or at least smoldering heavily, which seems really bad in my book if you're trying to make even reasonably good software.

> Those numbers don't really mean anything

IMO this is very dismissive. One example of probably many more, where software is used by millions and yet doesn't have this much being reported.

https://github.com/curl/curl/issues

curl isn't end-user software, and has a very small, well defined surface area.

curl is certainly a much more difficult problem to solve well, and does so without 30 million issues or whatever the CC codebase has

Curl is great, but I think you're vastly underestimating the surface area of CC.

It looks like it has more in: editor integrations, multiple guis and tuis, config variations, external systems (eg, git, mcps, curl-like requests?), statefulness (curl is "only" request response), inner runtimes (eg, sandbox per OS), sensitivity to its environment, possible side effects of its own execution, potential interaction combinations, etc.

Curl has hard system-level code requirements but it's design space feels more bounded and predictable to me.

This of course isn't an excuse for all bugs.

yes but there is a big difference between 100 sloppy features vs a handful of strongly and safely executed ones. it's not linear in effort.

> I think you're vastly underestimating the surface area of CC.

That is one of the reasons CC code is so low quality. The amount of extra unnecessary code just to implement it using React is by far the worst technical slop decision I've seen made by so called "software engineers" in a long time.

I dont't really know stats about such things, but I assume the adoption of CC has been insane compared to almost anything, and is also like two years old? Yeah, not surprised.

But also, if they all use 500 agents daily and latest models can "one shot" everything, one would think 5k issues are fixed in no time...

Havent used CC in some time, worked fine last time I tried.

> But also, if they all use 500 agents daily and latest models can "one shot" everything, one would think 5k issues are fixed in no time...

IMO, CC should be the poster child of what LLM coding should "feel" like. If things are so good why are there so many issues? Why can't they get a handle like other well-run projects? As you said, they have unlimited tokens so this project should be close to pristine as much as possible.

They're a small bootstrapped startup, give them some grace.

Small indie company, as the kids like to say.

Claude Code is also offered as an SDK, you can build custom (customized) harness on top of what essentially is Claude Code. https://code.claude.com/docs/en/agent-sdk/overview

Turns out he just used what he knew, TypeScript, and not what best solves the problem. He must have not had his "framework" then.

They have a UI which is pretty good! That explains why JS , they can run that in Electron and in TUI. I only use the UI now since it is do damn useful with its management of worktrees and multiple sessions, including ones running remotely via ssh devcontainers.

I would think that with agentic coding they’d be able to have a shared core and an interface native to the system, I.e. not react in the terminal and a swift or C# front end for the desktop app

They have a feature where they open a browser right in the app for Claude to use! They also use browser rendering for displaying various graphical formats. For once an Electron app that can actually justify shipping a whole browser in it.

I’ve rolled my own harness in Go. I have a rule to not let the production LoC exceed 50k lines. It is _very_ nice for my use cases. DeepSeek Flash V4.1 often performs at the level of GPT 5.6 Sol for non-orchestration tasks (programming and maths).

I add features and tweak it all the time: You just need to get comfortable with spending 3 hours in a chat session every 3 weeks to think hard about how to simplify whatever slop you didn’t simplify the last time you did this :)

Is it open? I'd love to take a look and perhaps steal some ideas :)

I am in the process of open sourcing it! Unfortunately, my employer requires a (hopefully) cursory legal review.

Replacing someone's words with a made up quote so you can dunk on them isn't how you display that you won an argument. I would ask that you engage in good faith with the other poster's ideas.

You cannot engage in good faith in a stupid argument. You can only point out it's stupid.

I’d recommend reading the late David Graeber’s “Debt” which does a good job illustrating exactly what you described, which is how currency is at its core an IOU

I cannot recommend reading this book enough.

So… does it follow that the very wealthy basically hold a massive IOU to the general public?

To expand the sibling comment. In principle the money would reflect IOUs the general public owes them. But also following this logic the way I see it is they accumulated their money was by getting more IOUs through: people employment (being able to command a big pool of work and get a cut of IOUs from it), arbitrage (exchanging lower value IOUs with higher value IOUs), or simply stealing (eg. siphoning government resources).

If by wealthy you mean people with a lot of money specifically, than no, the IOU is the other way around.

They hold IOUs from the general public.

That’s the implication I believe

I love this articulation.

Strong opinions tightly held

I try as hard as I can to have strong opinions, loosely held

For me, at least, despite being obviously powerful, I just can’t find anything useful to do with it beyond toys.

> most cars are in the ballpark

If you’re in the Bay, maybe, otherwise it’s $35-50k depending on how new

That's why I wrote "depending on where you are"