am.will

I've been testing Spark agents extensively, and I think I've found the roadmap to using them effectively. The secret isn't the model. It's how you control the context that gets passed to it. Everything I've learned swarming with Spark in Codex. 👇 x.com/LLMJunky/statu…

its not "a lot of extra metadata" i have tested this generally uses fewer than 2000 tokens. and clis/skills use tokens too. there is a difference but it is not a large one, and not enough to cause "context rot" clis are great, but they do not have bidirectional communication and are not nearly as secure. you cannot replace all mcps with clis. they both have their place and this idea that one is better than the other is just a fallacy. they are different. i'm totally good with you preferring CLIs, I do too, but i was addressing your comment about them causing 'context rot' they really are not a problem like that, especially with codex and anthropic's new 1M context window. this idea that mcp bad is parroted all across X and it just simply isn't true. much love

@LLMJunky @gener8ive @theo deleted that because i should have been more concise. mcp metadata not the entire tool. It’s okay if you like mcp! lol. but it’s a lot of extra metadata vs a cli for example that is stuffed into the context window

God I love Open Source. Integrated browser into T3Code Gonna add a terminal next.

@DamiDina @gustavonicot @OpenAIDevs only anthropic

@LLMJunky @gustavonicot @OpenAIDevs the important question - who has the data

WOW! I'm so excited about this. @OpenAIDevs said Codex was good at Computer Use, but I wasn't prepared for this. For the last two weeks I've been working on a Computer Use skill to work with Linux. And while I had some success, it was a pretty frustrating experience. That is...until the breakthrough. Using accessibility tools, Codex can now control my entire computer, not just the browser. There are limits to this, of course, but what a time to be alive. This Computer Use skill will unlock and entirely new set of automations, all powered by Codex. Demonstration below 👇

@JoshPurtell someone suggested it to me and it just worked. was happy

@LLMJunky Sounds like a familiar technique

Using accessibility tools 🤔

@daniel_mac8 @OpenAIDevs happy to share it with you. i'm gonna push it soon! just wanna test it out with more than just a browser

@LLMJunky @OpenAIDevs Looks very cool. Been running OpenClaw and Codex on a Linux desktop in my basement and would love to try it.

@DamiDina @gustavonicot @OpenAIDevs >i don't want to give it my password each time and want it to reliably do things like that imo as i said, you login ONCE and then its saved. just like you do in your normal browser. it is literally exactly what you're asking for

@LLMJunky @gustavonicot @OpenAIDevs i don't want to log in manually at all!

@shawnmvn @OpenAIDevs haha no they were first. they've had a computer use model since last year but its a bit...hard to use

@LLMJunky @OpenAIDevs you beat claude by 2 hours! :p

@dusty_irl @gener8ive @theo is that the goal? my goal isn't to reduce context rot. my goal is to build something. and that means giving it the right tools. and again, its not "stuffing" it into the context window because tools are only loaded when they are used. else, they are hidden.

@LLMJunky @gener8ive @theo if the goal is to reduce context rot, we have to be smart about what we are stuffing into the context window. that’s part of what makes pi so powerful

@ashiiirg @DevAdventur3s @OpenAIDevs yeah for web work for sure, but this isn't actually using screenshots with a11y afaik.

@LLMJunky @DevAdventur3s @OpenAIDevs The constant screengrabs to identify window positioning is a bit irritating though 🙄 I still prefer chrome tools MCP for web work

@gustavonicot @DamiDina @OpenAIDevs did it? claude code almost certainly drove MUCH more use than the app. i would need to see some data on that. no offense :) CLI isn't going anywhere. they share the same server so its trivial to support both

I agree. My point is slightly different. With skills, MCP, tools, and the Codex App server, you can already build almost anything. Capability is not the constraint. What I’m reacting to is how Claude shipped computer use. The app drove massive usage. If you ask me what could most impact adoption right now: 1.Native computer use 2.Simple remote control Everything else is already strong enough: models, coding, agentic capabilities. Time to market matters more than perfecting what is already good enough. Ship what drives usage. And from my perspective, that’s unlikely to be the CLI.

@DamiDina @gustavonicot @OpenAIDevs you dont have to. you can have it save sessions. you login manually, and then its saved

@LLMJunky @gustavonicot @OpenAIDevs i don't want to give it my password each time and want it to reliably do things like that imo

@johnennis @OpenAIDevs Thank you!

@LLMJunky @OpenAIDevs Man, that's amazing Great work

@joe_lgtm @OpenAIDevs You got to be careful with that. They'll wonder why you are suddenly productive lol

@LLMJunky @OpenAIDevs Codex, make it look like I’m doing work at my full time job while we build side projects. Make no mistakes.

@DevAdventur3s @OpenAIDevs Exactly so just have the agents create steps for you LOL

@LLMJunky @OpenAIDevs Yeh but why we have agents for right ? 😆

@dusty_irl @gener8ive @theo The token usage is very overblown, almost every agent has tool Discovery now to limit the number of tokens it uses. It's still more but not that crazy

@gener8ive @LLMJunky @theo mcp will cause context bloat. use pi extensions and create a runtime tool that fetches a called tool. So the runtime tool loader is loaded into context and then the llm requests the browser tool from available tools via the loader. that keeps it on-demand and will help with rot

@PachuDev01 @OpenAIDevs Well unfortunately we don't even have the Codex app officially just yet but we do have T3 code and I did Port the Codex app over so you do have options

@LLMJunky @OpenAIDevs I'm this 👌 close from moving to Linux, codex app feels so sloppy on windows, still using extension on vscode

@DamiDina @gustavonicot @OpenAIDevs You should be able to do that with just a regular browser use like agent browser

@LLMJunky @gustavonicot @OpenAIDevs i need computer use for clicking around random web ui like google cloud or random permissions not available via cli like creating a new instance of something

@danielrmay I was considering pivoting to a threadRipper but not anymore.

@LLMJunky what are u doing with it now

I had a few people laughing at me because I bought an RTX6000 Pro on credit without a plan. And you know what? They're right. It was FOMO. But there's a reason why. Check out this email I got from Newegg when I inquired about getting some DDR5 ECC memory. These shortages are real and not going anywhere, anytime soon. I was in a position where I could buy now and secure the best price possible, or risk watching GPUs and RAM continue to rise into unaffordability. For me, it felt like a now or never type of situation. Thankfully I have normal DDR5, board and CPU I can use. I have abandoned the threadripper path. Onwards!

@gustavonicot @DamiDina @OpenAIDevs You can do this in the Codex app too, it's just a skill. I don't agree that the CLI is a distraction because it uses the same app server under the hood. They share the same infrastructure. I do love the app though

@LLMJunky @DamiDina @OpenAIDevs I know many people are investing time in the CLI, but it feels like a distraction at this point. CLI is interesting, but I think the real shift is happening at their app layer with Codex App. That’s what’s driving Codex adoption.