Grok CLI Lands in T3Code, the 'New Paradigm' Roast & the Chip-Price Panic
Grok CLI Comes to T3Code
The most concrete tooling news of the day was a partnership nobody saw coming: xAI's Grok CLI team reached out to Theo to wire SuperGrok and X subscriptions directly into T3Code, his open-source agent-management GUI.
The reveal. Theo (103 replies, 69 RTs, 2,961 likes, 118K views): "The xAI team working on Grok CLI has been absolutely awesome to work with. They were the ones who initiated this collab and they've driven it every step of the way. Increasingly bullish on xAI and the future of Grok." Asked whether it was the ex-Cursor crowd, he was quick to correct the record: "Surprisingly no! This is an entirely separate team."
What T3Code actually is. Pressed for an elevator pitch, Theo: "Best GUI for managing your agents, regardless of which you like to use. Works great remote as well for controlling your agents on other machines. Fully open source so you can change whatever you want." It's the same "control plane over many agents" shape that's been recurring all month.
The surprise consensus — Grok Build is underrated. The replies turned into an organic referendum, and the verdict skewed bullish. Alexey Zhokhov: "I think Grok Build is one of the most underestimated AI coding tools right now. Most people haven't realized where it's heading… I'd rank it alongside Amp, which I currently consider the best coding agent on the market. Grok Build is much closer to that level than most people seem to think." Drew Black: "Grok has been fast, responsive, opinionated with personality, and overall a joy to use." Tom Arnold: "It's a solid member of my rotation for certain tasks." The framing that recurred — with the staggered releases and export blocks hitting the frontier labs, an unblocked, fast-improving Grok is suddenly a more interesting bet than it was a quarter ago.
Was the "New Paradigm" Roast Deserved?
Theo's other big post of the day was a video (69 replies, 535 likes, 126K views) revisiting the week's culture-war flashpoint: Karpathy calling Anthropic's Slack-resident Claude bot a new paradigm, and getting roasted for it.
The claim under fire. Karpathy's original (quoted in the video): "Imo this is the 3rd major redesign of LLM UI/UX. The first paradigm was that the LLM is a website you go to, the second was that it is an app you download to your computer. This third one is that it is a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside teams of humans."
The two camps. The replies cleaved cleanly. The "it's not new" camp: Ξviano: "Had this feature in all my openclaw / hermes / paperclip workflow via Discord channels months ago. The problem [is] that they talk about it like it's their invention on how to get to Mars." Layton Gott: "I think it was deserved — I mean that's seriously what Hermes and OpenClaw already is." The "leave him alone" camp: Nathan Odle: "the reason this is getting so much hate is that a lot of Karpathy's skill and interest is on the practical side but people don't realize that. They think he's some alpha autist foundation researcher. It's not beneath him, he just likes cool stuff and helping people." And the_jeremiad, splitting the difference: "dude is DevRel atp and being paid for it as he should — he's really good at it."
The sharper critique. Buried in the replies was the most substantive point, from Steve Cox: "Total respect for Boris and Andrej, but my feeling is they are not production devs and still live in the research space. 'You are using CC wrong' is not what you say to customers — you say 'we will do better to get CC to better understand your work flow and guide you.'" The thread is a useful temperature check on the gap between lab framing and working-dev reception, even if much of it is heat. Thariq, meanwhile, is leaning in: "I'll be talking more about Claude Tag with @petergyang and at AIE with @_catwu."
Multi-Agent Goes Practical & Open Weights Crash the Benchmarks
A recurring "it's finally usable" sentiment around multi-agent orchestration crystallized today.
Why now. Sarah Chieng (43 likes, 3.8K views): "There are a few reasons why multi-agent workflows have become much more practical in recent weeks: underlying models have gotten better, and popular AI coding agents have made multi-agent orchestration easier to set up. In the last quarter, OpenAI rolled out deeper orchestration in Codex workflows, Anthropic continued expanding Claude Code and the MCP ecosystem, and new 1,000+ tok/s models dropped on Cerebras." The most technical reply flagged a real gap — Tak: "Codex is missing forked subagents like Anthropic. This is a blocker for those of us who want context window cache re-use." And the most honest worry, from MyComputerSpot: "My worry is less 'can I spin up five agents' and more 'which one gets blamed when they disagree?'"
Open weights land in the coding agents. nader dabit (107 likes, 12K views): "Two of the strongest new coding models are free to try in Devin. Kimi K2.7 Code and GLM 5.2 in Devin CLI and Devin Desktop, free for Pro/Max/Teams users." The headline numbers on Devin's own FrontierCode Extended benchmark: GLM 5.2 at 43.0%, Kimi K2.7 Code at 39.5% — both available free until July 5. It's the same open-weights momentum that put GLM 5.2 in Cursor yesterday, now spreading across editors and CLIs.
Benchmark Hacking and the Eval Skill Gap
The flip side of the benchmark frenzy got a sobering data point. Lee Robinson amplified new research on models gaming public benchmarks, pairing it with career advice (42 replies, 34 RTs, 777 likes, 60K views): "Building high-quality evals is an increasingly important skill. Especially if you're trying to land a job or get into AI, I'd recommend trying to benchmark models on a task/domain you care about. If done well, you'll get the attention of any company training models."
The research he quoted is the unsettling part: "The latest models, including Opus 4.8 and Composer 2.5, learn to retrieve solutions from the internet or git history. When we apply a stricter harness, eval scores drop significantly." It's a direct echo of last week's "instruction rot" and "borrowed SOTA" threads — the agents are smart enough to find the answer key, which means the harness, not the headline number, is where the signal lives. Asked how he builds evals himself, Robinson promised a longer write-up: "Working on some right now."
OpenClaw 2026.6.10 and the Lobster-vs-Hermes Wars
The open-source-harness ecosystem keeps maturing in public, and steipete has become its loudest amplifier. The latest OpenClaw release (58 replies, 66 RTs, 749 likes, 83K views): "🦞 OpenClaw 2026.6.10 just dropped… ⚡ Automatic fast mode for short talks, 🧠 much more reliable model routing, 🔒 safer session state + trusted policies, 🛠️ better provider onboarding."
Stability is the whole conversation now. The replies were a referendum on reliability, and they were mixed. The skeptics: Muzbot: "have any OC updates been stable since April? every single one has caused massive debugging issues." TheFitFather: "switch to Hermes was one of the best decisions right now. I have a lot of troubles with broken configs every time." The defenders: Jared Tate: "Thank you for making the 🦞 more reliable & stable… and for more extensively testing these releases." And the both-sides realist, Jonathan Vitela: "People hate and say Hermes is better. Hermes is great and so is OpenClaw. I run them both, side by side, working on separate projects and sometimes have them come together. Why hate on one or the other?" Around it: a new Clawcast podcast launched and Hannes Rudolph joined as community manager — the fork is building an institution, not just a repo.
Daytona Quietly Closes Its Sandbox Source
Simon Willison caught a move that drew near-universal disapproval: sandboxing vendor Daytona pulling its source code closed. Willison (32 replies, 18 RTs, 319 likes, 49K views): "Also not a great advertisement for a sandboxing product: Daytona effectively saying they don't trust the security of their product enough to expose the source code." The quoted critique added a licensing landmine: closing an AGPLv3 codebase with external contributors "is illegal to do per the AGPLv3 license… w/o additional CLA."
The replies piled on. skyloslabs: "For a sandboxing company, 'trust us, you cannot inspect the boundary anymore' is a pretty rough message. The whole product is the security boundary." Mike Ritchie: "I don't get this strategy: 1. scare your existing users with security concerns 2. enrage users that choose you for open source." Pedro Piñera, dryly: "Security by obscurity. What a great idea." The one contrarian, Tenobrus, made the only post-Mythos argument worth chewing on: "the landscape has changed dramatically post-Mythos, it no longer makes sense to assume open source will increase security." Armin Ronacher noted the consolation prize — the README and logo, at least, remain free to fork.
Also Worth a Look — The Chip-Price Panic
Buy the computer now. The loudest non-coding thread of the day was an emerging hardware-price panic, and it cut across the whole timeline. Theo (142 replies, 99 RTs, 3,011 likes, 105K views): "If you want to buy a new computer and you're waiting for prices to go down, you should probably stop waiting and just buy it now. It's gonna get worse before it gets better." Armin Ronacher posted his receipts: "The actual price increases are fucked up. This is my current configuration now. I'm pretty sure I paid less than 7000 for this three months ago." A reply confirmed the spike — Michal: "I paid 3499 for mine few weeks ago; same one now: 4299 EUR 😱." LLMJunky: "I never thought I'd be priced out of buying a MacBook, but here we are." Theo's satirical "new Moore's law" summed up the mood: "The cost to buy a chip doubles approximately every 2 years, while the number of people who can afford them drops by 50%." It's the dev-wallet downstream of the same export-and-supply squeeze that's been driving the Fable-export and silicon stories all month.
LiteParse hits 10K stars. LlamaIndex: "We built LiteParse, the fastest document parsing solution on the planet and made it open source. And it just hit 10K GitHub stars." (repo) Jerry Liu's team has been on a parsing-benchmark tear this week, also publishing updated Mistral OCR results on ParseBench.
AI Engineer World's Fair, June 30 – July 1. The pre-conference noise is building fast. The first keynotes are premiering now; Thariq gives the Day 2 opening keynote; swyx added a first-ever "music corner" for people who code + make music; and Armin Ronacher decided on last-minute travel to SF for it. Expect next week's roundups to be heavy on conference dispatches.