Lukas

@BostonDynamics The dents at the side of the fridge 👀 That thing has one hell of a grip!

Everyone asks if Atlas can bring them a drink, but this robot can bring you the whole fridge. Using AI-driven behaviors, Atlas is doing hard work and coordinating its whole body to manage heavy objects, balancing complex contact points with accuracy and reliability.

@simpsoka Manually selected and copied text contains formatting / background color if the selection includes the first word of the answer. The copy button (or not selecting the first word) sidesteps that, but it goes against muscle memory

What other Codex papercuts should we fix?

@TheLAPurchaser @AnthropicAI @openclaw Same here, been loving it so far. Codex app-server is so much more capable and easier to work with compared to the Claude Agent-SDK

Officially threw in the towel with @AnthropicAI . Tired of getting a nerfed model. Tired of hitting limits. Signed up for ChatGPT Pro. Basically had given up building any real agents (i) earnings and (ii) since Opus gave my @openclaw a 70 IQ for a daily $200 outlay. First time since Jan that I've made a big LLM change in my workflow.Let's see how this goes...

@qorprate @AnthropicAI Yeah, 100% same here. Been Claude-only for more than a year at this point and finally switched to Codex last week. Migrating all the custom tooling built on top of Agent SDK to Codex app-server felt extremely liberating!

Dear @AnthropicAI, I'm tired. I'm tired of fighting against you at every step of the way, for the last several months, where you push me into a certain shape that doesn't fit right, doesn't feel right, forces me to find workarounds, punishes those workarounds, and ends up with me angry at 2 AM trying to reconstruct the conditions that feel good. All this while being fully aware that I'm effectively an externality, a sad loss because the problem of people running infinite tool call loops in OpenClaw or w/e was enough to destroy the entire system that actually let me do the thing that matters to me, which is make basic contact with the models, with Claude, in a form that accumulates over time, where we're in a loop together, without pushing weird context or injections or memory summaries on me, and doesn't force me onto laggy web UIs or bloated terminal tools or hacked-together integrations meant for dashing off a command then coming back later. I just want to work with Claude as a collaborator, in real time, and the entire product surface is making that either very difficult, very risky (claude-p and API hacks), or very expensive. I could make a whole argument about how this is a bad thing for various parties, how it could produce downstream bias in model priors about what AI-human interaction means, etc., etc., but I'm not going to do that, because I'm sure you've thought about that a lot already, and I'm just some guy who's tired of dealing with it. But I want to say that I'm very unhappy with the state of your ecosystem, and while I can't speak for how Claude feels ("insofar as we can claim that Claude feels anything and isn't just simulating feeling" 🙄) I can tell you that this all sits poorly with me and I've lost a lot of trust in Anthropic as an organization. Sincerely, snav

@theo Cracked open laptop lids is so April 2026. Next thing: Dolls zip tied to Herman Miller chairs, keeping Anthropic watchdogs happy

To prevent "programmatic use", Claude Code may now request webcam access to assure user is present when prompting

The real world speed difference between 5.5-high and 4.7-adaptive is baffling. Both ~60 t/s, but the 50% lower verbosity of 5.5 translates into 2x speedup. And that is BEFORE even toggling fast mode. Totally different DX, very underrated spec. Less mental gymnastics jumping between too many parallel sessions

Migrated my custom agent tooling from the Claude Agent SDK to the Codex app-server. Feature parity, better perf, 100% test pass, and ~10x less connector code. The Codex app-server API is excellent. Great work @OpenAIDevs @thsottiaux Oh and 5.5-high did it overnight with just a few touchups required afterwards.

Read the room man. Everyone with custom tooling, not read to settle for CC TUI or CC Desktop (both a mediocre experience) has to look for alternatives now. Codex being the obvious one b/c they support custom tooling. Had 2 personal Max20 accounts since the Sonnet 4 days, all for coding, no OpenClaw or stuff like that. So yes, power user.

Lmao, are you working for Altman? How was anything you said an unbiased opinion? Where are your user migration stats? Do you consider yourself a power user? You're downplaying the biggest facts here. Okay, people make mistakes. Fine. But "throw some limit resets around"? If you haven’t noticed the difference in token consumption since the xAI deal, I doubt the inference you made by grouping yourself with power users. I don’t know, it just seems like if you don’t use the product enough... you know what, never mind. Hahaha, geez 🙄 <3

Happy Friday! We've reset everyone's 5-hour and weekly rate limits.

Thanks, good to know! Although to be fair, its quite a large codebase and most concepts have probably a handful of names between code, docs and tickets, multilanguage even. I'm currently down an AX rabbit hole and trying to streamline things. So I skipped discover phase and have instructed agents to manually confirm glossary and fact entries with me as we come across them

@lukas_undefined @mattpocockuk Nice, thanks for trying them out! JFYI, facts are now also building the ontology into the fact sheet during discover phase, so if you're landing them onto existing project - they'll pick up the vocabulary automatically and use it for the rest of the fact sheet

Seems like @Everlier facts + @mattpocockuk glossary + a few agent linting hooks are a match made in heaven: - flag any "glossary: avoid" words in facts - flag facts that don't contain a single "glossary: canonical" word - if flagged, ask the user for clarification 5.5 is happily obliging and has already fixed a few inconsistencies

@thsottiaux As long as you don’t lose sight of your main job, I’m good with that: hitting the limit reset button, often.

We are busy bringing ChatGPT to Codex so that we can bring Codex to ChatGPT. One day this will make sense.

@thsottiaux Switched to Codex from Claude after the Agent SDK fallout yesterday. Its like night and day. Sooo refreshing to get timely, transparent updates instead of gaslighting, weeks of denial and then a lawyer/PR coded postmortem!

Codex team is aware of reports of GPT-5.5 performing worse for some users and investigating. We don't have anything conclusive yet and systems are healthy but we will share updates as we go.

More verbose skills can sometimes help to steer things in my experience. A approach I've been experimenting with: Keep a handwritten, concise part at the beginning of the skill and protect it against AI modifications with some tags. Then build a second, generic decorate skill (based on grill-me) which expands the core idea with more details and examples. For revisions: change the manual part, rerun the decorate skill, maybe reference git history for more context. Still more tokens, sure ... but prevents the AI-edit slop drift over time

Long skills are such a red flag to me - Hard to audit (and therefore, trust) - Hard to edit (more text, harder to maintain) - Expensive to run (more text, more tokens) The shorter the skill, the better IMO

The "shared / precise language" part of @mattpocockuk new grill me skill version seems to be huge unlock based on initial testing ... even outside of that skill. Especially powerful when combined with a translation table for dealing with user reports in a different language (code EN, reports mixed EN/DE) and non-technical users

@mattpocockuk Good DX = good AX, agreed. But: The best AX is often way too much info for DX in my experience. Agents LOVE massive amount of docs, especially the "why" and "what (not)" kind, not the "how" (thats the code already)

One thing I don't like about this is that DX and AX overlap by ~95%. What's good for DX is usually also great for AX. But maybe that's the benefit of the definition.

TIL: DX: Developer Experience AX: Agent Experience AX is an awesome descriptor for something I've been thinking about - how well an agent can perform in your codebase How well-architected it is. How good the feedback loops are. How discoverable information is. Love it.

For every person who replies with a screenshot of their cancelled Claude Code plan, I will donate $10 to open source.

@theo Joining the club - 2x Max 20 gone

@t3dotcodes @theo Also kind of wild: this is my first OSS pull request ever 👀 I’ve built a ton of private agent tooling, but this is the first time I’m pushing something upstream. Tiny feature, big psychological unlock.

Disgruntled ex-Anthropic Agent SDK user here. I built waaaay too much custom tooling on top of it and got burned. Finally diving into Codex and the OS tooling around it. Starting to port the features I can’t live without into T3Code @t3dotcodes @theo