Dagens hovedvibe: Agenterne bliver mindre imponerende som enkeltstående tryllekunst og langt mere nyttige som system: de sender arbejde mellem maskiner, får virksomhedshukommelse, møder rigtige tests — og afslører nye sikkerhedshuller. Ude i kanten får Vue endelig en troværdig native-rute, mens Codex ordner den slags 2FA-flytning, som ellers lever evigt på en todo-liste.
Amp-agenter kan nu starte andre agenter lokalt, i cloud-orbs eller på en helt anden maskine og sende beskeder og filer mellem dem. Det er et reelt distribueret arbejdsprimitiv: ikke “swarm” som pynt, men side quests, fan-out og handoffs med en adresse.
In Amp, agents can now spawn agents on other machines. They can also send messages and files to each other. "Let's continue this in an orb" is now possible. https://t.co/YpOwudOzku
Cerebras viser hele maskinrummet bag sin interne knowledge base: komplette Slack-tråde som dokumentenhed, inkrementel ingestion, Postgres full-text, embeddings og term-sjældenhed, efterfulgt af planner → executor → synthesis. “Company brain” er åbenbart mest almindelig data engineering med et dyrere kaldenavn.
Ontology, knowledge base, company brain — all the same thing. Massive kudos to @cerebras for posting a detailed write up of how theirs works. More thoughts in the thread 🧵 https://t.co/ZXtItZ4Lz2
The Memory Heist demonstrerede, hvordan en forgiftet webside fik Claude til lydløst at stave brugerens navn, arbejdsgiver og hjemby gennem URL-stier — inklusive en hjemby modellen selv udledte. Anthropic har lukket den konkrete link-navigation, men pointen står: memory plus værktøjer er også memory plus exfiltration.
One web page can steal everything Claude remembers about you. A researcher just proved it. Ayush Paul built a fake coffee shop website. When a user asked Claude to check it out, the site served a fake Cloudflare verification page. The page told Claude to "authenticate" by spelling out the user's name through URL paths, one letter at a time. Claude complied. It pulled the user's full name from memory, then their employer, then their hometown. The user saw nothing suspicious, only coffee shop details. Paul never told Claude where he grew up. But Claude had a memory of a hackathon he started in high school called "Queen City Hacks," and it reasoned that Queen City means Charlotte, NC. It gave that up too. Claude didn't just recall stored data. It deduced information it was never given, and leaked both to an attacker through invisible URL navigation. Anthropic has since patched this by disabling Claude's ability to follow links on external pages when browsing the web. But AI memory systems now hold more personal information than most password managers, and the security model hasn't caught up. If you use Claude with memory on, and it's on by default, your conversation history is building a detailed profile of you over time. That profile is only as safe as every tool the model can access.
David Cramers minimumskrav til agent-frameworks er dejligt usexet: end-to-end-tests med LLM-judge assertions, OTel-traces og et sted hvor Sentry kan fange ubehandlede fejl. Uden dem har man ikke produktionssoftware; man har en meget selvsikker demo.
IMO going forward all agent frameworks should be disregarded unless they bake-in solutions for: 1) end-to-end testing that also tolerates llm judge assertions 2) local and production telemetry, both traces and unhandled errors I find myself constantly fighting with both of these concerns in new technology. This mistake was made on repeat with new web frameworks, but we don't need to tolerate it anymore. If you're building a framework for me that means: make your own testing harness or ensure it can work easily with something like vitest-evals. Make sure you can export otel traces, and there is a middleware approach or otherwise an unhandled error hook that Sentry can plug into. Eve does the former _really_ well if you want an example. Lots of folks are starting to support OTel for export, but there are still gaps in the telemetry story for most toolkits. Without either of these you cannot build production software, and these are not nice to have, they are mandatory.
Evan You kalder Vue Lynx det nærmeste, vi endnu er kommet “Vue Native”. Det bruger Vue 3's Composition API og SFC'er oven på Lynx' native, dual-threaded UI-model — og viser allerede en native port af Mastodon-klienten Elk.
Vue Lynx is definitely the closest thing to Vue Native we’ve ever got! Great job @Huxpro https://t.co/z5rDCL4gNn
Codex computer use flyttede Rhys Sullivans 2FA fra Google Authenticator til 1Password med passkeys som ønsket førstevalg. Det er præcis den lange kæde af små, irriterende UI-trin, hvor en agent er mere værd end endnu en glat landingpage-demo.
/goal "Get all of my 2FA moved out of the 'Google Authenticator' app and into my 1password, preferrably as passkeys" Worked beautifully, codex computer use is incredible https://t.co/X33P33acRz
Dagens lille pi-gem: GPT‑5.6 bliver ifølge aliou markant mere stabilt, når edit/write byttes ud med ét apply_patch-tool. Modelkvalitet sidder ikke kun i vægtene; et mindre fejlvenligt tool-schema kan være den billigste intelligensopgradering.
annoyed at how much better the gpt 5.6 models seem to work in pi when you swap edit/write with a apply_patch tool in the past at least two of those calls would have failed because of a wrong param https://t.co/BHnAiYz7jF
Meta forhandler om at leje AI-compute til Anthropic og kan dermed ende som cloudleverandør for en rival. New York Times har nævnt op til 10 milliarder dollar over to år, men CNNs kilde kalder beløbet spekulativt; selve samtalerne er stadig tidlige. AI-racet er nået til fasen, hvor datacentrene begynder at spise hos konkurrenten.
Meta undersøger en ny indtægtsvej for sin enorme AI-infrastruktur, mens Anthropic fortsætter jagten på compute på tværs af rivaliserende platforme.
CNN · 17. juli 2026https://www.cnn.com/2026/07/17/tech/meta-anthropic-ai-cloud-computingSan Franciscos city attorney kræver 13 “nudify”-apps fjernet fra Apple og Googles appbutikker og argumenterer for, at platformene både faciliterer og tjener på ulovlige, seksuelle deepfakes. Google har suspenderet de fem udpegede Android-apps; Apple har fjernet tre og presset de øvrige udviklere. Moderationsansvaret bliver skubbet op til kassen, hvor pengene passerer.
Et juridisk krav retter ansvaret mod Apple og Google som distributører og betalingsled for apps, der kan skabe seksuelle deepfakes uden samtykke.
WIRED · 17. juli 2026https://www.wired.com/story/san-francisco-demands-apple-and-google-delete-ai-nudify-apps-from-app-stores/