X / Twitter
Swyx of the Latent Space podcast and AI Engineer conference
Swyx proposed a concept called Loopcraft: he believes the central game of the next century is who can stack loops most effectively. Early in each stage, knowing when to drill down one loop for reliability is valuable; as models improve, knowing how to move up one loop for greater impact may become more valuable. He put it bluntly: people who do not understand this should not complain when they lose to those who do. In another post, he explained his main motivation for building a vibe-coding platform himself: he likes Vercel, Cloudflare, and Netlify, but none truly closes the loop. None puts you back on track when something goes wrong or proactively pings you when a system fails—and systems always fail. Each project requires repeating webmaster-style infrastructure setup, with npx posthog wizard here and npx arize skills there; he just wants to absorb all of it into One Thing once and for all. He also congratulated the Ona team on joining OpenAI and linked to their talk, suggesting it contains clues to Codex’s next direction. Overall, his focus in recent days has been consistent: in the agent era, the advantage lies with whoever closes the feedback and infrastructure loops.
https://x.com/swyx/status/2065307558198567206
https://x.com/swyx/status/2065264832056889711
https://x.com/swyx/status/2065176231453282777
OpenAI Codex and ChatGPT team member Thibault Sottiaux
Thibault confirmed the combination of Codex and Ona with the brief caption “Codex 🤟 Ona.” He said he was incredibly excited to build the future with Johannes and his team. Together with Swyx’s congratulations that day on Ona joining OpenAI, this marks another clear team reinforcement for OpenAI in agentic coding. For people watching Codex’s roadmap, it is a signal worth noting: the Ona team’s capabilities will directly enter Codex’s evolution.
https://x.com/thsottiaux/status/2065193272952422852
Replit CEO Amjad Masad
Amjad has recently been sharing his experiences since Fable arrived in Replit. He said it was the first time he had vibe-coded with zero frustration and entered a complete flow state, to the point of finishing his backlog of ideas and starting to run out of things to build. His conclusion was that vibe coding no longer needs higher IQ, only cheaper and faster models, and then it is done. On costs, he praised the Replit Agent team for making Fable affordable, saying its fewer mistakes actually make it cheaper overall. He also showed what starting a company on Replit looks like: a single canvas containing your web app, mobile app, marketing materials, and App Store assets, where clicking any item lets you continue building, editing, or generating new things. Taken together, his message is that Replit wants to put the whole process of “starting a company” into an agent-native workspace.
https://x.com/amasad/status/2065236013627351551
https://x.com/amasad/status/2065259509082411233
https://x.com/amasad/status/2065241626436583860
Vercel CEO Guillermo Rauch
Guillermo reposted a practical Vercel + Shopify example: developer foda used v0 and Cursor to build a fully custom Next.js headless storefront that processed more than 500 orders in 2 minutes. His conclusion was long-term optimism about the web: anyone can now complete the full sequence of dream, build, ship, and sell. He also briefly teased Vercel + Grok without elaborating. Together, the posts emphasize that Vercel is connecting both the commerce side, Shopify, and the model side, Grok, positioning itself in the middle of the path from AI website building to selling products.
https://x.com/rauchg/status/2065116986678624419
https://x.com/rauchg/status/2065118448947216681
Box CEO Aaron Levie
Aaron announced a study Box had just conducted, surveying 1,640 IT leaders across the US, Japan, and Europe about agentic AI adoption. Its most striking finding was that companies with the highest AI adoption also planned to hire the most people. He acknowledged that the data allows multiple interpretations and contains confounding variables, but argued the logic is intuitive: companies with the largest productivity gains have both the willingness and the means to reinvest the benefits in their businesses and keep growing. He directly challenged the “AI destroys jobs” narrative, which assumes that the total amount of work companies want to do is fixed. What is actually happening, he argued, is that AI makes companies want to launch more engineering projects, sell to more customers, and automate more processes, all of which create more work for people. For those building enterprise AI, this survey of 1,640 people provides an argument they can cite.
https://x.com/levie/status/2065287110744297809
Y Combinator CEO Garry Tan
Garry used two posts to rebut an article rejecting the concept of “gifted children.” Author Katie Arnold-Ratliff cited a 35-year longitudinal study in which only 12.3% of 677 gifted children achieved “eminence,” defined as becoming full professors, Fortune 500 executives, or federal judges, using this to argue the classification is a lie. Garry said the finding actually refutes the author: the base rate for such rare achievements in the general population is far lower, close to zero rather than a double-digit percentage, so 12.3% indicates that the selection mechanism works extremely well. He also noted that Community Notes had already identified a problem with the author’s own understanding of “gifted,” and called for an end to authors using lies to look cool by signaling allegiance within their circles. In another post, he shared California Governor Gavin Newsom’s visit to YC for a garryslist event, where they discussed little tech, YIMBY, stopping the asset seizure tax, and sustaining California’s innovation wave while ensuring all Californians prosper. In recent days he has been both sparring in public discourse and laying groundwork for innovation policy in business and political circles.
https://x.com/garrytan/status/2065314389196959813
https://x.com/garrytan/status/2065313198237180238
https://x.com/garrytan/status/2065298785463579053
OpenClaw creator Peter Steinberger
Peter shared a concrete development in OpenClaw’s security hardening. Media format conversions previously required shelling out to ffmpeg, expanding the attack surface. In the next release, this can be done with wasm, with similar performance for their use cases. It is a straightforward example of reducing surface risk: replacing external process calls with wasm execution inside a sandbox substantially narrows the attack surface. He also showed everyday collaboration involving Chris using Codex to submit PRs, and remarked that writing Mac apps remains difficult. For people building agent infrastructure, the move from ffmpeg to wasm is worth studying.
https://x.com/steipete/status/2064999763397980286
https://x.com/steipete/status/2065176989359808636
https://x.com/steipete/status/2065132980398444945
Every CEO Dan Shipper
Dan shared a less successful experience with Fable. He set it up with a large project and left it running on its own, then returned an hour later to discover that it had triggered safeguards after 10 minutes and fallen back to 4.8. His response was to go straight back to Codex. This is firsthand feedback from a heavy user of a new model: stability during extended autonomous runs matters more than peak capability in determining which tool they stay with.
https://x.com/danshipper/status/2065269582961737957
Official blogs
Anthropic Engineering: How we contain Claude across products
Anthropic’s engineering team systematically reviewed the isolation architectures and pitfalls of three agent products: claude.ai, Claude Code, and Claude Cowork. Its central position is to first establish hard isolation at the environment layer through sandboxes, VMs, and egress controls, then add probabilistic defenses at the model layer. Key figures include: Claude Code users had approved around 93% of permission prompts, making human oversight ineffective through approval fatigue; OS-level sandboxing reduced permission prompts by 84%; and Claude Opus 4.7 had an attack success rate of around 0.1% for a single attempt on Gray Swan’s Agent Red Teaming benchmark, rising to around 5% to 6% after 100 adaptive attempts. The article candidly disclosed real incidents: internal red-team phishing persuaded employees to paste a malicious prompt, after which Claude exfiltrated AWS credentials in 24 of 25 attempts; and an attacker used their own Anthropic API key to send user files to their account through the allowlisted api.anthropic.com. The latter produced an important conceptual shift: “A domain allowlist is not a destination filter but a capability grant.” Another recurring lesson was that “what you build yourself is the weakest part”: proven primitives such as gVisor, seccomp, and hypervisors did not fail; the failures were all in custom proxies and configuration-loading logic. For teams building agent products, this is one of the most information-rich recent accounts of practical security engineering.
https://www.anthropic.com/engineering/how-we-contain-claude
Podcasts
Training Data — Google DeepMind's Logan Kilpatrick: Why the Model Eats the Harness
Key takeaway: The agent harnesses everyone is rushing to build themselves today will be “eaten” by models within 12 months, and the alpha will move elsewhere.
Logan Kilpatrick leads Google AI Studio and Gemini API and is a central figure in Google’s developer ecosystem. Recorded after Google IO, this conversation unusually brings together a full account of Google’s internal agent strategy, its real position in the coding competition, and its world-model direction.
The first focus was Antigravity’s actual positioning. It is not just an IDE but an ecosystem, including a core IDE, a web agent experience, a CLI, and an SDK. More importantly, the same agent harness is becoming the new through line across Google products, powering agent features in Search, the Gemini app, Cloud, and AI Studio. Historically, Gemini models connected more than 50 Google products; now that connection is the harness itself. The base harness is around 80% shared, with specialization for use cases such as vibe coding or 7x24 consumer agents.
Second was his candid assessment of the coding competition. The host said developer friends were roughly split between Claude and Codex, with few using Gemini, and he did not evade the issue. His explanation had three layers. When Gemini 3 launched last December, the narrative was still “Google has won,” but the agentic coding wave immediately rewrote it after the holidays, showing how fast things change. The fundamental reason for struggling to build a long-horizon SWE-level coding model is that you need such a product of your own to drive the flywheel, which explains the Windsurf deal and Antigravity’s creation. And outsiders have overlooked the pretraining window: 3.5 Flash exceeded every previous Pro model’s coding capabilities through post-training alone, while the results of large-scale pretraining are still on the way.
Third was the argument in the title. Two years ago, a “model” was a set of weights; today, a model is an entire system around the weights, including tool calls, hosted search, code execution, and container runtimes. Scaffolding always runs a few steps ahead of models, then gets absorbed. He therefore predicts that today’s ubiquitous claim that “the alpha is in the harness” will probably no longer hold in 12 months, and suggests the industry create a harness benchmark to measure how well different models adapt to different harnesses.
Several data points are worth remembering: 350,000 Android apps have been built on AI Studio since last week, many of them personal apps that would never otherwise have been created. Finance, much of it crypto-related, accounts for around 20% of app categories, while games used to account for 20%. He predicts that within this year, ordinary people will be able to vibe-code genuinely fun games. The bottleneck is not model quality, but that people who understand games have yet to build the right scaffolding.
Discussing GDM’s culture, he quoted one of his favorite lines from Silicon Valley: “We can’t let anyone outdo us at making the world a better place.” He said this is the whole industry’s current condition: everyone is competing to make the world better. Framed that way, it seems rather funny, and the endeavor is fundamentally not zero-sum.
For founders building tools, the most practical lesson from this episode is not to base their moat on harness engineering, but on focus in a vertical and understanding customers. “Focus is a startup’s superpower,” while the window for harness advantages is closing.