Shopify drops React Native for Swift and Kotlin, credits coding agents
Shopify is moving its mobile apps from React Native back to Swift and Kotlin, and its engineering blog says coding agents are why.
Shopify moved its apps to React Native six years ago so it wouldn't have to write everything twice. Now coding agents have cut the cost of writing it twice, so it's going back to native. Anthropic says a Chinese rival secretly answered its customers with Claude, while they thought they were using Kimi. OpenAI turned its own Codex coding agent into a rentable API. Cognition's new coding model edges Anthropic's Fable 5.1 on one agent benchmark. On a newer one, it scores under half of Fable's mark. Claude Code's creator says AI-written production code needs a higher bar than a human's. The same week, a critic let OpenAI's GPT-6 Astra run for 35 hours and got "absolutely nothing of value." And Claude Code's desktop app now pops any pane into its own window.
• Shopify drops React Native, credits coding agents
• Anthropic says Moonshot secretly served Claude as Kimi
• OpenAI opens its Codex harness as a public API
• Cognition's new model wins one benchmark, bombs another
• Claude Code's creator and a critic clash on trust
• Claude Code's desktop app pops panes into their own windows
Mobile agents
Shopify drops React Native for Swift and Kotlin, credits coding agents
Shopify is rebuilding its mobile apps in Swift and Kotlin, six years after moving them to React Native specifically to avoid writing everything twice. Its engineering blog says coding agents changed the math: they "could implement a feature on Android using the iOS version as a reference, and vice versa," cutting the cost of keeping two platforms in sync. The post names no speed-up number. Shopify's React Native libraries get handed off or wound down: it's talking to several companies about taking over FlashList, and Restyle is being archived, kept working through the end of 2026. It's the second Shopify story in two days that runs through coding agents. Yesterday: Tailwind, whose creator had said AI cut its revenue. Today: its own apps.
Mustafa Ali
Threat intelligence
Anthropic says Moonshot secretly served Claude to Kimi users
Anthropic's latest threat report accuses Moonshot AI of secretly routing Kimi users' requests to Claude and training on the responses — almost 300,000 requests in one ten-day stretch, through 5,380 fraudulent accounts. One case: an engineer at a Chinese state-owned enterprise used what they thought was Kimi to build an internal tool. Internal code and live credentials went to Anthropic. Anthropic doesn't know if Moonshot told its customers. Moonshot had not responded on X as of Friday morning, US time.
We're publishing our most detailed threat intelligence report to date.
— Anthropic (@AnthropicAI) September 10, 2026
It covers how people tried to misuse Claude—for cyberattacks, influence operations, surveillance, biology, and building weapons—and how we found and stopped them.
We disrupted every operation in the report,…
Agent infrastructure
OpenAI opens its Codex coding harness as a public API
OpenAI is renting out the Codex harness that runs its own coding agent, as a public API in beta, with no fee beyond the tokens and tools used. It manages sessions, orchestration, context compaction and recovery; your app supplies the tools and picks where the sandboxed agent runs, MCP servers included. HN's top reply is the right skepticism: nobody has settled what an "agent product" looks like yet.
Benchmarks
Cognition's new model wins one benchmark, bombs another
Cognition's SWE-2 scores 92.8% on Terminal-Bench 2.1, edging Anthropic's Fable 5.1 at 91.4% — and 27.3% on the newer Terminal-Bench 4, where Fable 5.1 scores 55.8%. The model is post-trained from Moonshot's Kimi K3. Anthropic's report, out the same day, doesn't mention K3. Every number above is Cognition's own internal eval, not reproduced independently. It ships today in Cognition's Devin Desktop and CLI.
Introducing SWE-2, our closest model yet to the frontier.
— Cognition (@cognition) September 10, 2026
On leading evals, it scores on par with recent frontier models – at up to 70% lower cost.
We scaled RL to multiple trillions of parameters, with a refined recipe that pushes the Pareto curve on both capabilities & cost. pic.twitter.com/FjBxrikExs
Trust
Claude Code's creator and a critic clash on trusting AI-written code
Claude Code's creator Boris Cherny drew a line: throwaway prototypes can be a black box, but "production code written by Claude should have a higher bar than if it was written by a human." Days earlier, developer Armin Ronacher tested the opposite, letting OpenAI's GPT-6 Astra run unattended for 35 hours on one prompt: a net 75k lines across 79 commits, "around 1200 USD in raw API costs," and "absolutely nothing of value." Ronacher's verdict: "I do not manage to trust this model much."
Hey ████,
— Boris Cherny (@bcherny) September 11, 2026
I think there is room for both.
1. Prototypes and other throw-away code can be treated as totally black box. If you’re going to throw it away anyway, and if the blast radius of it breaking is low, it doesn’t need to be perfect.
2. Production code written by Claude…
Tools
Claude Code's desktop app pops any pane into its own window
Claude Code's desktop app now lets you pop any pane into its own window. Drag the diff or terminal to a second screen, and Claude keeps working in the main window. Dock it back whenever you want. Splitting panes inside one window already worked; the separate window is the new part. Cherny also turned /diff into a persistent pane you can scroll and click, updating in real time.
You can pop out any pane in the Claude Code desktop app into its own window.
— ClaudeDevs (@ClaudeDevs) September 10, 2026
Drag the diff or terminal to a second screen while Claude keeps working in the main window, then dock it back whenever you want. You can also run sessions side-by-side or stacked. pic.twitter.com/5BF38zm4ky
Also worth your time
- Claude Code v2.1.268 — fixes deny and ask permission rules on symlinked directories not applying when a path was given by its real location, a Read or Edit deny rule not applying when an
env -Corevalcommand sat on the same line, and WebFetch hanging on a server that never finishes a response (it now fails after 300 seconds). - A Tell HN says OpenAI keeps re-enabling its "allow training" setting — "I've reset this more than once and the last time I made a careful note of when I did it and to my surprise I found it re-enabled when I checked just now." One user's report; the thread splits on whether it happened to others.
- GitHub's AI Scan for pull requests gets REST APIs — organization and repository endpoints to turn code scanning's AI Scan on across repos, in public preview; separately, MAI-Code-1-Flash is deprecated across all Copilot experiences as of September 10.
- OpenAI is retiring GPT-5.3-Codex-Spark next week — "Next week we'll be retiring GPT-5.3-Codex-Spark. Can you believe we shipped a model named as such!!"
Know someone who'd want this in their inbox? Forward it — that's how this grows. And if we got something wrong, or you think we buried the real story today, hit reply. A person reads every one.
The New Way is written with AI. It gathers the day's stories, checks them against their sources and drafts every summary. A person decides what runs and reviews every issue before we hit send.
