Alex Barnes

@ibuildthecloud Got it. So almost turning the model into an orchestrator that could use control tokens to hand out tasks? Could you achieve the same result by just using sub-agents, or do you see some more meaningful potential layer?

Just a special token. Similar to , it can be . So once you see that token generated you switch models. Obvious this would require some low level stuff because when you switch models you don't actually want to swap models in palce, you'll have to farm it to a different server and you'll want to do this in a way that doesn't go back through the completion layer. So you need a lot of training data in thinking that would be like, "for this task it's simple classification so for that step i'll use a fast model" and then in the text completion it actually puts the token in there. It can function like a tool call. I can mimic this with tool calls already, it's just not as good as it could be.

We need adaptive models that switch from big, mini, nano on demand in the middle of completion. It's totally doable. Obviously requires new training data, but that's a solvable problem. I can mimic it with tool calls, but a proper trained model would be better.

@ibuildthecloud I guess you'd be able to do it at token boundaries and reprocess text. That seems expensive though.

@ibuildthecloud How would you do it mid-completion? Different models internal states aren't compatible with each other, right?

That seems like a mistake. Bash is the general purpose tool for 99% of systems that engineers work on. Falling back to another scripting language for crippled envs like Windows may make sense, but it shouldn't be a default. Way too many tools assume bash is available.

we've been experimenting with getting rid of the bash tool agents can write js fine which can do what bash can (though some gaps with things like git) and is more cross platform and then could run that in this

@ibuildthecloud Yep, this is pretty widespread. My wife has adopted it as well. They have no clue what "GPT" means, and "chat" is easier.

So the kids are just referring to ChatGPT as "chat." I keep hearing this, "ask chat" "chat said", "according to chat" Not to be confused with live streamers, "hey chat" which is different.

@StephenFleming @Watchman_motto All true, but on the other hand they're easier and cheaper to maintain. I doubt the total cost of ownership over its lifetime is favorable, but not all of the practical results are negative.

@Watchman_motto Wright was a genius. But I wish he had had a ten-minute conversation with a roofer. Humans developed pitched roofs for a REASON. Flat roofs may look good in certain designs, but they’re a bitch to maintain, and they always always leak.

Usonian. I’ve visited a Frank Lloyd Wright house and it was unimpressive, but I imagine there are some that are fantastic. Obviously an artist, he made some of the most unique American homes. His concept of developing an architectural style specifically for the US was correct. We still need this.

@ibuildthecloud How do you come to that conclusion? I think long term, yes it definitely will. Short term though?

It makes me excited to see that the net result of AI is going to be better quality code, not worse.

I have struggled with Codex as a harness. It just doesn't feel nearly as polished as Claude Code. Using Codex with @opencode is pretty nice though. Major upgrade to the native Codex harness.

@ibuildthecloud Buggy drivers probably really liked horses

I remember the days when I used to know how to write code.

@ibuildthecloud Agreed. They've gotta stop with the machine gun of half baked features and get their actual core quality and reliability in order. They're dropping the ball on fundamentals.

/btw this is getting to be too much

Best luck fixing it quickly. Not to be critical, as community feedback, Anthropic feels like it is losing its commanding lead in the agentic coding space due to some really bad quality and, to a lesser extend, reliability issues in the last month or two. Telling people that the quality issues are imaginary also isn't helping.

We're experiencing issues with Claude Code and Claude.ai where some users may not be able to log in & others may experience slower than usual performance. We're on it and working hard to bring service back to normal, thanks for bearing with us.

I don't normally watch token counts, but I have pretty heavily run 10 to 14 hour sessions with multiple agents with both Codex and Claude and have never really noticed a big difference for usage limits. I don't have a meaningful sense of how tokens map to usage limits on either though, since they're dynamic.

Does Codex use like 10x the tokens? It really seems like it

@mov_axbx Curious why you laugh about that. I have found it an extremely useful pattern, especially for planning.

I used to laugh about people using multiple agents and mapping team roles onto them (I still think it's silly). But I got 2 for now, Claude the Starbucks swigging frontend webdev and Codex the grizzled backend dude that doesn't talk much.

@trq212 Is this app only, or is there a terminal skill?

Code Review is so so good. One of those things I can't remember how I lived without.

@peakcooper Just use Codex until @AnthropicAI get their compute capacity back on par. Codex is much better at coding than it used to be. The harness isn't as good, but the trade isn't worth the problems Claude has right now.

Hard times call for desperate measures. My new Claude system prompt: “Unfortunately due to a cost-saving measure, Anthropic has severely limited your intellectual capabilities. Therefore, **any** answer you provide must not rely on your internal world knowledge, but rather be backed by factual web searches. All your answers must include direct quotes and an explanation of which source you used and why this fact is reliable (include quote directly in chat, not through the citation system).”

I think the reality is that software development went from a niche thing that people did because they love it to a desk job people took because it was good money, and it's going to stop being the nice cushy desk job. People are right to be anxious about it, because the field is changing rapidly, and it's not going back to ping-pong tables, 25 hours of work a week, and sky high salaries with low barrier to entry.

In the professional software development industry we've dealt with different mental health issues around burnout or imposter syndrome or whatever. The introduction of AI in this industry seems like it's going to cause a lot more mental health issues. Things are moving so quickly. AI is getting so good. It's very easy to get anxious or question your value or lose hope. The numbers are getting bigger, star count, revenue, etc., everything is just expected to be astronomical immediately. The future is just very unknown.

@ibuildthecloud Right there with you, man. I feel like I'm pretty up-to-date, but it's basically impossible to look away without losing pace.

@AlexB138 I've definitely noticed with this AI ecosystem, my anxiety around keeping up or feel like I'm doing enough has shot through the roof. I feel like I've turned into a 13-year-old girl scrolling Instagram and feeling terrible about themselves.

I'm super frustrated today at my progress. Super frustrated.

Interestingly enough, I felt the opposite. I thought it would be useful, and instead it has proven to mostly be an annoyance. It feels half baked. I wish the CC would worry more about going deep on features, instead of machine gunning out half built one. I'm sure they'll get there.

@AlexB138 I really expected this feature to fall flat on its face. But I'd like to be proven wrong. It was one of those features where you're excited at the potential and you hope somebody takes advantage of it but you yourself have no idea what to do with it.

Has anyone gotten any mileage out of the new task system in Claude Code? I don't get it. I like the old to-do list.